Medical quality management method and system based on big data and artificial intelligence

Abstract

The invention relates to the technical field of medicine quality management, in particular to a medicine quality management method and system based on big data and artificial intelligence, and the method comprises the steps: obtaining the quality influence parameters of a medicine product in the production, transportation and storage processes through a multi-modal data collection device; noise is removed through data cleaning, data dimension reduction is carried out through principal component analysis and an auto-encoder, the analysis efficiency is improved, and the efficiency problem of mass data processing of a traditional method is solved. Time sequence analysis is carried out through the long-short-term memory network, abnormal nodes are recognized in time, manual sampling inspection is replaced, and accurate early warning is achieved. The graph neural network constructs a path tracking model, analyzes the relevance between nodes, traces the quality problem, and reckons the batch of affected drugs. Based on a traceability result, an XGBoost model generates quality risk scores of all links by analyzing historical and real-time data and sets a threshold value, and quality risks are early warned in real time.

Description

Technical Field

[0001] The present invention relates to the technical field of pharmaceutical quality management, and in particular to a pharmaceutical quality management method and system based on big data and artificial intelligence. Background Art

[0002] The quality of pharmaceutical products is affected by complex factors in multiple links during production, transportation and storage, such as production environment conditions, equipment operating status, transportation and storage conditions, etc. These factors largely determine the stability and safety of drug quality. In the production process, environmental factors such as temperature, humidity, and air pressure will directly affect the stability of drug ingredients. Especially for some environmentally sensitive drugs, their quality changes may occur in small environmental fluctuations. At the same time, fluctuations in the operating status of production equipment such as temperature control systems and pressure systems may also lead to quality differences between drug batches. The transportation link faces more complex influencing factors. The transportation of drugs needs to be carried out under specified environmental conditions. Fluctuations in transportation temperature and humidity will affect the chemical properties of drugs. Especially in long-distance transportation, drugs are easily affected by extreme weather and changes in transportation conditions, which in turn leads to uncontrollable fluctuations in drug quality. Changes in storage conditions such as fluctuations in temperature and humidity will also affect the shelf life and safety of drugs.

[0003] Traditional quality management methods rely on manual testing and regular sampling, which has limitations and cannot cover the entire production and transportation process, and cannot provide real-time warnings for potential quality problems. In addition, traditional methods are inefficient in processing massive, multi-dimensional parameters affecting pharmaceutical quality, and cannot effectively analyze the correlation between different data sources, making it difficult to trace quality problems, affecting the efficiency of tracing drug quality problems. Summary of the invention

[0004] To solve the above problems, the present invention provides a pharmaceutical quality management method and system based on big data and artificial intelligence. By acquiring a large amount of parameter data affecting the quality of pharmaceutical products from production to transportation, abnormal problems are identified, traced and risks predicted through artificial intelligence, thereby solving the problem of low traceability efficiency in traditional quality management methods.

[0005] To achieve the above object, the technical solution adopted by the present invention is: A pharmaceutical quality management method based on big data and artificial intelligence comprises the following steps: S1. Obtaining quality-influencing parameters of pharmaceutical products from production to transportation through multimodal data acquisition equipment, wherein the quality-influencing parameters include production environment parameters, production equipment working parameters, pharmaceutical ingredient data, pharmaceutical batch information, transportation and storage status, and pharmaceutical quality test results; S2. Clean the data of the quality-influencing parameters and perform data standardization, and perform dimensionality reduction processing on the standardized quality-influencing parameters through principal component analysis and autoencoder; S3. Perform time series analysis on the quality-influencing parameters after dimension reduction based on long short-term memory networks to identify abnormal nodes in the production and transportation processes; S4. Establish a path tracing model for the entire life cycle of pharmaceutical products based on graph neural networks. The path tracing model learns the correlation between nodes, traces the source of quality problems based on abnormal nodes, and calculates the affected drug batches and IDs to obtain traceability analysis results. S5. Based on the results of traceability analysis, the quality risk prediction of real-time quality influencing parameters is carried out through the XGBoost model.

[0006] Furthermore, the production environment parameters include the temperature, humidity and air pressure of the production environment, the production equipment working parameters include the operating parameters of the equipment in each link of the pharmaceutical product production, the drug ingredient data include the drug ingredient raw materials and the corresponding batches, the drug batch information includes the production batch and production date of the drug, the transportation and storage status includes the transportation route point, transportation and storage temperature, transportation and storage humidity and transportation and storage air pressure, and the drug quality inspection results include whether there is any abnormality in the drug quality.

[0007] Further, the S2 comprises the following steps: Apply semantic segmentation model to classify quality-influencing parameters and assign type labels; Apply dynamic data cleaning rules to remove noise and outliers for quality-influencing parameters of different types of tags; Setting a hierarchical normalization function to perform nonlinear normalization on production environment parameters and transportation state parameters; The quality influencing parameters of production, transportation and storage dimensions are clustered by principal component analysis and K-means clustering algorithm to obtain the central features. The core feature data is obtained by performing nonlinear mapping and dimensionality reduction on the central features through the autoencoder.

[0008] Further, the S3 includes: A long short-term memory network model is constructed based on the core feature data. The core feature data generated by the real-time quality influencing parameters are compared with the core feature data generated by the historical quality influencing parameters based on the long short-term memory network model to calculate the probability of quality anomaly and generate abnormal nodes in the production, transportation or storage process.

[0009] Furthermore, the path tracing model for the entire life cycle of pharmaceutical products is established based on the graph neural network according to the abnormal nodes, including the following steps: Modeling the links in the life cycle of a pharmaceutical product as nodes in a graph neural network model, wherein the links in the life cycle of a pharmaceutical product include production equipment, transportation tools, and storage facilities; Based on the quality impact parameters collected at each node, a data association matrix between nodes is generated; Based on the graph neural network, deep learning is performed on the data association matrix between nodes. By recursively processing the input and output data of the nodes, the data flow between each node is extracted, the influence weights between the nodes are calculated, the path weight matrix is ​​generated, and the path tracking model is constructed through the path weight matrix.

[0010] Furthermore, the path tracing model is used to learn the correlation between nodes, trace the source of the quality problem, and infer the affected drug batches and IDs to obtain the traceability analysis results, including the following steps: Based on the path weight matrix of the path tracing model, starting from the abnormal node, the influence weight of each node on its upstream node is recursively calculated, and the propagation path of the abnormal node is identified by comparing historical data with real-time data; Based on the abnormal propagation path and the drug batch information of the node, batch tracking is performed to calculate the drug batches and IDs affected by the abnormality, and the traceability analysis results are summarized and generated.

[0011] Further, the S5 comprises the following steps: Based on the traceability analysis results, the corresponding quality-influencing parameters of the drugs affected by the anomaly are extracted, and the contribution of each parameter to the quality risk is calculated using the gradient boosting algorithm; Based on the training parameters selected by contribution, the XGBoost model is used to gradually train the corresponding training parameters of pharmaceutical products at production, transportation and storage nodes to generate quality risk scores for each node; Set risk thresholds through quality risk scoring and issue real-time risk warnings based on the risk thresholds.

[0012] Furthermore, the formula of the XGBoost model is as follows: ; in, Score quality risks; is the output of the kth regression tree; is the weight of the corresponding quality influencing parameter; Quality influence parameter used for regression trees.

[0013] A pharmaceutical quality management system based on big data and artificial intelligence, applied to any of the above-mentioned pharmaceutical quality management methods based on big data and artificial intelligence, comprising a big data acquisition module, a data cleaning module, an anomaly identification module, a traceability analysis module and a quality risk prediction module connected in sequence; A big data acquisition module is used to obtain quality-influencing parameters of pharmaceutical products from production to transportation through multimodal data acquisition equipment, wherein the quality-influencing parameters include production environment parameters, production equipment working parameters, pharmaceutical ingredient data, pharmaceutical batch information, and transportation and storage status; The data cleaning module is used to clean the quality-influencing parameters and perform data standardization, and perform dimensionality reduction processing on the standardized quality-influencing parameters through principal component analysis and autoencoder; The anomaly identification module is used to perform time series analysis on the quality-influencing parameters after dimensionality reduction based on the long short-term memory network and identify abnormal nodes in the production and transportation process; The traceability analysis module is used to establish a path tracing model for the entire life cycle of pharmaceutical products based on a graph neural network. The path tracing model learns the correlation between nodes, traces the source of quality problems based on abnormal nodes, and infers the affected drug batches and IDs to obtain traceability analysis results. The quality risk prediction module is used to predict the quality risk of real-time quality influencing parameters based on the results of traceability analysis through the XGBoost model.

[0014] The beneficial effect of the present invention is that the present invention obtains the quality influencing parameters of pharmaceutical products in the production, transportation and storage links through multimodal data acquisition equipment, including production environment parameters (such as temperature, humidity, air pressure, etc.), production equipment working status, drug ingredient data, drug batch information, transportation and storage status and drug quality test results. This real-time data acquisition method makes up for the lag of traditional reliance on manual detection and sampling, can achieve real-time monitoring of the entire life cycle, effectively capture small fluctuations in the production and transportation process, and especially timely monitor and feedback the quality of environmentally sensitive drugs. Noise is removed by data cleaning technology to ensure the integrity and accuracy of the data, and the standardized quality influencing parameters are reduced in dimension by principal component analysis (PCA) and autoencoders. Through this data dimensionality reduction processing, data redundancy is greatly reduced, and the key features that have the greatest impact on drug quality are retained, thereby improving the efficiency of data analysis and solving the problem of low efficiency of traditional methods in processing massive data. By performing time series analysis on the quality influencing parameters after dimensionality reduction through a long short-term memory network (LSTM), abnormal nodes can be identified in a timely manner. The LSTM model has the ability to process time series data and can detect potential quality anomalies based on the comparison between historical data and real-time data. This can not only effectively replace the traditional manual sampling method, but also further improve the accuracy of anomaly identification through intelligent algorithms, so as to achieve timely warning. The path tracking model of the entire life cycle of pharmaceutical products is constructed through the graph neural network (GNN), and the correlation between each node is learned to help analyze and locate the source of quality problems. Through the correlation analysis of nodes such as production equipment, transportation tools, and storage facilities, the specific links and equipment that cause the problem can be accurately traced back to the abnormal nodes. At the same time, the batch information can also be used to infer the batches and IDs of drugs that may be affected, which greatly improves the traceability efficiency of quality problems and solves the problem of difficult efficient traceability in traditional methods. Based on the traceability analysis results, this solution also uses the XGBoost model to predict the risk of real-time quality influencing parameters. By analyzing historical data and real-time data through the XGBoost model, the quality risk scores of each production, transportation, and storage link can be generated, and thresholds can be set according to the risk scores to trigger the real-time warning mechanism to ensure that risks can be predicted before problems occur. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a step flow chart of a pharmaceutical quality management method based on big data and artificial intelligence in the present invention.

[0016] Figure 2 It is a step flow chart of step S5 in the present invention. DETAILED DESCRIPTION

[0017] See also Figure 1-Figure 2As shown, the present invention relates to a pharmaceutical quality management method based on big data and artificial intelligence, comprising the following steps: S1. Obtaining quality-influencing parameters of pharmaceutical products from production to transportation through multimodal data acquisition equipment, wherein the quality-influencing parameters include production environment parameters, production equipment working parameters, pharmaceutical ingredient data, pharmaceutical batch information, transportation and storage status, and pharmaceutical quality test results; S2. Clean the data of the quality-influencing parameters and perform data standardization, and perform dimensionality reduction processing on the standardized quality-influencing parameters through principal component analysis and autoencoder; S3. Perform time series analysis on the quality-influencing parameters after dimension reduction based on long short-term memory networks to identify abnormal nodes in the production and transportation processes; S4. Establish a path tracing model for the entire life cycle of pharmaceutical products based on graph neural networks. The path tracing model learns the correlation between nodes, traces the source of quality problems based on abnormal nodes, and calculates the affected drug batches and IDs to obtain traceability analysis results. S5. Based on the results of traceability analysis, the quality risk prediction of real-time quality influencing parameters is carried out through the XGBoost model.

[0018] In some embodiments, first, the quality-influencing parameters of pharmaceutical products during production, transportation and storage are acquired in real time through a multimodal data acquisition device. The acquired parameters include production environment conditions (such as temperature, humidity, air pressure, etc.), which are crucial to the stability of drug ingredients. In addition, it is also necessary to acquire the working state parameters of the production equipment (such as the operating temperature, pressure, speed, etc. of the equipment), which can reflect the state of the equipment operation and its potential impact on the quality of the drug. For example, fluctuations in the temperature control system may directly lead to quality differences between batches of drugs. In addition, drug ingredient data, batch information, and transportation and storage status also need to be acquired synchronously to ensure comprehensive monitoring of all links. In the transportation link, it is particularly important to pay attention to the real-time changes in transportation temperature and humidity, because these environmental factors may have irreversible effects on the chemical properties of the drug. Next, the acquired quality-influencing parameters are cleaned. The main purpose of this step is to remove noise, abnormal data, and incomplete data to ensure the integrity and consistency of the data. After data cleaning, the standardization processing stage is entered. In this process, principal component analysis (PCA) is used to reduce the dimension of the standardized data to reduce data redundancy. Dimensionality reduction not only improves computational efficiency, but also retains the most representative features, ensuring effective analysis of key quality influencing factors. Subsequently, the autoencoder is used to further extract features from the data, compressing the high-dimensional raw data into lower-dimensional core feature data, which is finally used for subsequent time series analysis. Based on the data after dimensionality reduction, the long short-term memory network (LSTM) is used for time series analysis. LSTM is a deep learning algorithm suitable for processing time series data. It can capture the time dependency and historical fluctuation patterns in the production and transportation of pharmaceutical products. By comparing historical data with current real-time data, LSTM can identify potential abnormal nodes in the production and transportation process. For example, if the operating temperature or pressure of the equipment is abnormal during a certain period of time, LSTM will promptly identify abnormal fluctuations based on the trend of historical data and mark the relevant abnormal nodes, thereby helping enterprises to take corrective measures in a timely manner. Based on the abnormal nodes identified above, a graph neural network (GNN) is used to establish a path tracking model for the entire life cycle of pharmaceutical products. The model analyzes and tracks the transmission path of abnormal nodes by learning the data correlation between each node. Specifically, the graph neural network can learn the influence weight of each node from the associated data between multiple nodes such as production equipment, transportation tools, storage facilities, etc., and trace the specific links that may cause quality problems based on the propagation path of abnormal nodes. Through this path tracing model, the potential source of quality problems can be effectively identified, and the affected drug batches and IDs can be inferred. This process not only improves the traceability efficiency of quality problems, but also provides a strong basis for subsequent quality control and production optimization. Finally, based on the results of the traceability analysis, the XGBoost model is used to predict quality risks for real-time quality influencing parameters.The XGBoost model is an efficient and accurate integrated learning method that can accurately predict future quality risks by training historical data. In this embodiment, the XGBoost model generates a quality risk score for each node based on historical traceability data and current real-time monitoring data. By setting a risk threshold, when the risk score of a node exceeds the predetermined threshold, the system will automatically trigger a risk warning mechanism, thereby notifying relevant personnel in advance to intervene.

[0019] Furthermore, the production environment parameters include the temperature, humidity and air pressure of the production environment, the production equipment working parameters include the operating parameters of the equipment in each link of the pharmaceutical product production, the drug ingredient data include the drug ingredient raw materials and the corresponding batches, the drug batch information includes the production batch and production date of the drug, the transportation and storage status includes the transportation route point, transportation and storage temperature, transportation and storage humidity and transportation and storage air pressure, and the drug quality inspection results include whether there is any abnormality in the drug quality.

[0020] Specifically, multimodal data acquisition equipment is used to obtain quality-influencing parameters during the production, transportation and storage of pharmaceutical products, ensuring that the full life cycle quality monitoring of pharmaceutical products can be fully covered. Among them, the production environment parameters specifically include temperature, humidity and air pressure that are monitored in real time during the production process. For example, for drugs that need to be produced under constant temperature and humidity conditions, the system will continuously collect temperature and humidity data in the workshop to ensure that they are within the set range. If there is an over-standard situation (such as temperature fluctuation exceeding ±2°C), the system will record and feedback in time. In addition, air pressure parameters are used to monitor air flow changes in the production workshop. Especially in the production of drugs that are highly sensitive to environmental conditions, air pressure fluctuations may cause changes in product stability. The working parameters of production equipment cover the equipment operation data involved in all aspects of production. For multiple key equipment in a production line, such as temperature control devices, compressors, mixers, etc., the system will collect key data such as operating temperature, pressure, and speed in real time. For example, in the process of mixing drug components, the speed and pressure of the equipment are crucial to the uniformity and quality of the drug. If the speed fluctuates abnormally or the pressure is unstable, it may cause insufficient or excessive mixing of the components, thereby affecting the final quality of the drug. By accurately monitoring these working parameters, the system can detect potential problems in equipment operation in a timely manner and issue early warnings. Drug ingredient data includes the ingredient raw materials used in drugs and their corresponding batch information. Each batch of raw materials will be quality tested before entering the production link, and its source and batch number will be recorded to ensure that different batches of raw materials can be traced. By recording the raw material batch information and correlating it with the subsequent drug quality test results, the system can trace back to the specific raw material supply source when quality problems occur. For example, if a component deviation is found in the quality test of a batch of drugs, the system can trace back to the corresponding raw material batch through the component data, and determine whether the batch of raw materials has quality abnormalities through historical data analysis. Drug batch information includes the production date and production batch number of each batch of drugs. This information will be recorded in detail throughout the life cycle of the drug, including its status in production, transportation, storage and other links. This information is particularly critical for quality traceability. When a batch of drugs has problems in subsequent quality tests, the system can quickly locate the relevant production records through the batch number and production date, and conduct multi-dimensional analysis in combination with the environment and equipment parameters to determine the specific cause of the quality problem. During transportation and storage, the system will continuously monitor data such as transportation route points, transportation temperature, transportation humidity, and air pressure. For example, during long-distance transportation, drugs need to be kept in a certain low temperature and low humidity environment. The system installs sensors in the transport vehicles to record temperature and humidity changes in real time. If the temperature exceeds the set standard range (such as exceeding 2-8°C), the system will generate an abnormal alarm record and notify relevant personnel to handle it.In addition, records of transportation route points can help track each key node of the drug during transportation, so that the complete transportation path of the drug can be traced when quality problems occur. Drug quality test results are evaluated in a variety of ways, including online monitoring systems and regular laboratory analysis. The system will correlate and compare various types of data monitored in real time with the laboratory's quality test results to determine whether there are any abnormalities in the drug quality. For example, a batch of drugs was found to have chemical composition that did not meet expectations during laboratory testing. The system can combine the temperature and humidity data of transportation and storage to analyze whether the chemical changes of the drugs were caused by environmental fluctuations during transportation.

[0021] Further, the S2 comprises the following steps: Apply semantic segmentation model to classify quality-influencing parameters and assign type labels; Apply dynamic data cleaning rules to remove noise and outliers for quality-influencing parameters of different types of tags; Setting a hierarchical normalization function to perform nonlinear normalization on production environment parameters and transportation state parameters; The quality influencing parameters of production, transportation and storage dimensions are clustered by principal component analysis and K-means clustering algorithm to obtain the central features. The core feature data is obtained by performing nonlinear mapping and dimensionality reduction on the central features through the autoencoder.

[0022] In some embodiments, first, the system applies a semantic segmentation model to classify quality-influencing parameters. The semantic segmentation model extracts features from the input multimodal data and divides data from different sources by type. Specifically, the model performs layer-by-layer convolution operations on the original data to extract low-level and high-level features. Low-level features mainly reflect local information such as noise and outliers, while high-level features can capture global information such as temperature, humidity, and equipment working status. Based on semantic segmentation, the system assigns a type label to each type of data. For example, production environment parameters (temperature, humidity, etc.) will be marked as environmental, production equipment operating parameters (equipment temperature, speed, etc.) will be marked as equipment, and drug ingredient data will be associated with raw material sources and batch information, and will be input as component classification data for the next step of processing. After the data classification is completed, the system applies dynamic data cleaning rules to clean quality-influencing parameters of different types of labels. For production environment parameters, the cleaning rules automatically remove abnormal points according to preset environmental thresholds, such as values ​​of temperature and humidity that exceed the normal range. The dynamic cleaning rules use a sliding window algorithm to determine whether a data point is abnormal by calculating the local mean of time series data. If the value of a data point exceeds the historical average range in the sliding window, the system will mark the data point as an outlier and remove it. For the working parameters of production equipment, the system uses a statistical noise filtering method to smooth the obvious noise recorded by the equipment sensor to ensure the accuracy of the data. The parameter cleaning of the transportation and storage status depends on the trend of historical data. The fluctuations of transportation temperature and humidity are analyzed through the difference detection algorithm to eliminate irrelevant anomalies caused by sudden changes in the external environment. After completing the data cleaning, the system enters the hierarchical standardization processing stage. This step applies different standardization strategies to different types of data. In the processing of production environment parameters, the system uses a linear normalization method to standardize data such as temperature and humidity to the range of 0 to 1. This process calculates the minimum and maximum values ​​of each data and scales the data according to the linear formula (x-min) / (max-min) to ensure that environmental data from different sources are subsequently processed at the same scale. For temperature and humidity data in the transportation and storage status, the system uses a nonlinear normalization method, such as the Logistic normalization function, which is suitable for processing data with a large range of fluctuations, can better handle extreme situations, and prevent the data from losing key fluctuation information during the normalization process. Next, the system performs feature clustering on the normalized quality-influencing parameters using principal component analysis (PCA) and K-means clustering algorithms. Principal component analysis calculates the covariance matrix of the quality-influencing parameters and extracts the principal components with the greatest explanatory power in the data. In this process, the system first performs eigenvalue decomposition on the covariance matrix to obtain eigenvectors and eigenvalues, and selects the first few principal components with the largest variance as the basic features for clustering.These principal components represent the directions with the largest variance in the original data set, which can effectively reduce the data dimension and avoid noise and redundant information in high-dimensional data. After PCA processing, the system uses the K-means clustering algorithm to cluster the principal component data. The K-means algorithm randomly selects the initial cluster center, assigns each data point to the cluster center closest to it, and repeatedly iteratively adjusts the position of the cluster center until the difference in the data within the cluster is minimized. Through this clustering method, the system can summarize the multidimensional quality influencing parameters in production, transportation and storage into several representative central features. Finally, the system applies the autoencoder to perform nonlinear mapping and dimensionality reduction on the clustering results. The autoencoder consists of a set of neural networks and is divided into two parts: encoder and decoder. In the encoder part, the system compresses the input high-dimensional features layer by layer through nonlinear activation functions (such as ReLU or Sigmoid), maps the original high-dimensional data into a low-dimensional latent space, and obtains the core feature data. In the decoder part, the system attempts to restore the original data from the core features in the low-dimensional space to ensure that important information is not lost during the dimensionality reduction process. For example, when processing production equipment parameters, the autoencoder compresses multiple parameters of the equipment (such as speed, temperature, and pressure) into a low-dimensional feature vector by capturing the operating mode of the equipment, and retains the information on the impact of these features on drug quality. The core features after dimensionality reduction are not only convenient for subsequent model processing, but also can remove redundant secondary information, ensuring the accuracy and efficiency of data analysis.

[0023] Further, the S3 includes: A long short-term memory network model is constructed based on the core feature data. The core feature data generated by the real-time quality influencing parameters are compared with the core feature data generated by the historical quality influencing parameters based on the long short-term memory network model to calculate the probability of quality anomaly and generate abnormal nodes in the production, transportation or storage process.

[0024] It should be noted that, first, a long short-term memory network model (LSTM) is constructed based on the core feature data generated in step S2. As a deep learning model suitable for processing time series data, LSTM can effectively capture long-term dependencies and short-term fluctuations. In particular, in the production and transportation of pharmaceutical products, quality parameters often have time dependence and periodicity. Through its special structure, LSTM uses the mechanism of input gate, forget gate and output gate to determine which information needs to be retained and which information needs to be forgotten at each moment. For example, in the process of monitoring drug production, environmental parameters such as temperature and humidity may fluctuate greatly in the short term, while some equipment status parameters have long-term stable characteristics. By learning the time variation law of these parameters, LSTM can automatically adjust their weights and capture important time series features. Next, the LSTM model compares the core feature data of the quality-influencing parameters collected in real time with the historical data. In this step, the LSTM model first processes the real-time quality parameters in the current production or transportation, extracts its time series features, and generates feature vectors. At the same time, the model extracts feature vectors under the same time dimension from the historical data and calculates the similarity of the two vectors. Specifically, LSTM generates a time series similarity score by calculating the time series difference between real-time data and historical data. For example, if the production temperature, humidity and other parameters of a batch are significantly different from the data under normal historical production conditions, the model will determine the possibility of abnormality in the batch. Then, the model will calculate the probability of quality abnormality based on these feature differences. LSTM predicts whether there is abnormal fluctuation in the current time series through the activation value generated by its output layer, combined with the comparison results of real-time and historical data. The specific calculation method is: the model calculates the error between the actual observation value and the predicted value through the loss function, and updates the weight of the model based on the error size using the back propagation algorithm, and finally outputs a probability value of quality abnormality. Assuming that the equipment pressure of a batch of drugs fluctuates during the production process, the LSTM model will determine whether the fluctuation is an abnormal fluctuation based on the normal production parameters in the historical data, and output a probability value indicating the possibility of quality abnormality under the current production conditions. When the probability of abnormality exceeds the preset threshold, the LSTM model will mark the time point as an abnormal node and record the relevant parameters. Abnormal nodes can reflect critical moments that may affect the quality of drugs during production, transportation or storage. For example, if during long-distance transportation, the LSTM model detects that the temperature continues to fluctuate beyond the expected range for a period of time, the system will mark that time point as an abnormal node and generate an abnormal report based on the relevant temperature and humidity data. This abnormal node can not only provide an important basis for tracing the source of quality problems, but also help production and management personnel to adjust production or transportation conditions in a timely manner to avoid the expansion of potential quality risks.

[0025] Furthermore, the path tracing model for the entire life cycle of pharmaceutical products is established based on the graph neural network according to the abnormal nodes, including the following steps: Modeling the links in the life cycle of a pharmaceutical product as nodes in a graph neural network model, wherein the links in the life cycle of a pharmaceutical product include production equipment, transportation tools, and storage facilities; Based on the quality impact parameters collected at each node, a data association matrix between nodes is generated; Based on the graph neural network, deep learning is performed on the data association matrix between nodes. By recursively processing the input and output data of the nodes, the data flow between each node is extracted, the influence weights between the nodes are calculated, the path weight matrix is ​​generated, and the path tracking model is constructed through the path weight matrix.

[0026] In some embodiments, first, the life cycle of a pharmaceutical product is modeled as a node in a graph neural network. Specifically, each node represents a key device or facility in the production, transportation, and storage process. For example, production equipment such as mixers and compressors are regarded as nodes, and transportation tools such as cold chain transport vehicles and storage facilities such as cold storage are also modeled as independent nodes. The core of the graph neural network is to model these nodes and construct a relationship network between nodes so that data can flow and transfer between different nodes. Node feature data includes physical parameters such as temperature, pressure, and speed of the equipment, as well as environmental data such as transportation temperature and humidity. These multi-dimensional parameters are used as input to form the initial state vector of the node. Next, a data association matrix is ​​generated based on the quality influencing parameters collected by each node. Each element of the matrix represents the degree of mutual influence between two nodes. By constructing this association matrix, the system can convert a multidimensional data set into a matrix form that can be processed by deep learning. The algorithm for generating the association matrix is ​​based on calculating the similarity between node feature vectors, and uses, for example, cosine similarity or Euclidean distance to quantify the strength of association between different nodes. In the generated association matrix, a larger value indicates a stronger association between nodes, and vice versa. This association matrix is ​​the basis for subsequent learning of the graph neural network, which can reflect the potential interactive relationship between each node in the production and transportation links. Subsequently, the system conducts deep learning on the generated association matrix through the graph neural network (GNN). The recursive message passing mechanism of the graph neural network plays a key role in this link. Specifically, the GNN model first propagates the initial information to each node. Through the message passing mechanism, each node transmits its own status information (such as equipment operating status, environmental parameters, etc.) to adjacent nodes, and receives information from other nodes at the same time. In this multi-round iterative message passing process, the graph neural network aggregates information from adjacent nodes and updates the state vector of the node. After each round of iteration, the feature vector of the node not only contains its own information, but also integrates data from other nodes. The specific calculation process depends on the convolution aggregation operation, which is expressed as: ; in, is the feature vector of node v after the k+1th iteration; AGG is an aggregation function used to aggregate the features of neighboring nodes; is the neighbor node of node v; is the set of neighbor nodes of node v; is the weight matrix used for feature transformation; is a nonlinear activation function (such as ReLU). This recursive calculation process allows the graph neural network to capture the complex interactions and data flows between nodes, and then extract the global features of the system. Based on these recursively learned features, the system further calculates the influence weights to measure the actual impact of each node on its neighboring nodes. The influence weights are generated by calculating the weight matrix. The algorithm determines the contribution of each node to its neighboring nodes based on the correlation between the strength of association between nodes and the changes in node features. The influence weight matrix reflects the quality impact path of each link in the entire pharmaceutical product life cycle. For example, when the temperature fluctuation of transportation equipment A affects the humidity of subsequent storage equipment B, the system evaluates the specific impact of temperature fluctuations on humidity changes through the calculation of the weight matrix. Finally, based on these weights, the system generates a path weight matrix and uses this matrix to construct a path tracking model in the pharmaceutical product life cycle. The model can accurately identify abnormal nodes and their propagation paths from production to transportation to storage. For example, if a node is marked as an abnormal node, the system will analyze how the abnormality is transmitted from the production stage to the transportation or storage stage based on the path weight matrix. This path tracking can help companies quickly locate the source of abnormal problems and optimize quality control and management processes.

[0027] Furthermore, the path tracing model is used to learn the correlation between nodes, trace the source of the quality problem, and infer the affected drug batches and IDs to obtain the traceability analysis results, including the following steps: Based on the path weight matrix of the path tracing model, starting from the abnormal node, the influence weight of each node on its upstream node is recursively calculated, and the propagation path of the abnormal node is identified by comparing historical data with real-time data; Based on the abnormal propagation path and the drug batch information of the node, batch tracking is performed to calculate the drug batches and IDs affected by the abnormality, and the traceability analysis results are summarized and generated.

[0028] Specifically, first, through the path tracing model, the system starts the traceability process according to the identification of the abnormal node and the path weight matrix. Assuming that the temperature control device (node ​​A) on the production line has abnormal temperature fluctuations during a certain period of time, the system will treat the device as an abnormal node and use the graph neural network (GNN) model to track the path of abnormal propagation. The path weight matrix plays a core role in this process. The matrix defines the influence weight of each node on other nodes. The calculation of path weight is based on the aforementioned recursive message passing mechanism. The state change of node A (such as temperature fluctuation) will affect its neighboring nodes (such as mixing device B) through message passing, and then propagate backward to other downstream nodes along the production process. Specifically, the system starts from the abnormal node and recursively calculates the influence weight of each node on its upstream node to identify how the abnormality propagates in the network. Through the recursive calculation of the graph neural network, the model can analyze the impact of the abnormality of device A on subsequent devices layer by layer and evaluate its propagation path in the data chain. In the recursive process, each node receives information from its upstream node and updates its own state based on the information. Assume that the abnormal temperature fluctuation of device A is transmitted to mixing device B, and the speed of mixing device B may be affected by the fluctuation, which further affects the drug packaging quality of downstream device C. Through this iterative recursion, the graph neural network accurately tracks the transmission process of the abnormality in the entire production chain and records the degree of influence of each node. In the recursive process, the system also combines historical data with real-time data to further confirm the propagation path of the abnormal node. For example, device A may have experienced similar temperature fluctuations in the historical production cycle. The system compares historical data to evaluate whether this abnormality is similar to past abnormalities, and analyzes the severity and scope of the abnormality based on this. The combination of real-time data and historical data enables the system to conduct a more accurate traceability analysis of the abnormal transmission process, ensuring that the identified abnormal propagation path is valid and has historical basis. Next, based on the abnormal propagation path, the system associates the drug batch information of the node with the path for batch tracking. During the production process, each production link (node) will record the batch information of the drug, such as the batch number recorded by the production equipment and the batch status of the transportation tool. The system determines which batches of drugs may be affected by the anomaly by matching the nodes in the anomaly propagation path with its batch data. Assuming that during the time period when the anomaly occurred in production equipment A, it affected a batch of drugs (batch number Batch123) that was being produced, and the batch of drugs was subsequently produced through mixing equipment B and packaging equipment C, the system will infer based on the results of path tracking that the anomaly affected Batch123 and may have also affected subsequent parts of the batch (such as transportation or storage). In order to ensure the accuracy of batch tracking, the system will further combine the historical records of batch tracking to screen other batches that may be affected.For example, if the abnormality of equipment A has similar records in the production of the past batches, the system will analyze its correlation with the current abnormal batch and infer the potential affected batches (such as Batch124). In addition, if the abnormality propagates to the subsequent storage link, the system can also infer the drug ID that passed through the path during the abnormality through the status information of the transportation and storage equipment, and trace it according to its circulation process. Finally, the system generates traceability analysis results based on the above analysis, summarizes the drug batches and IDs affected by the abnormal nodes, and provides detailed abnormal propagation paths and impact assessments. For example, the traceability analysis report will clearly point out which nodes the Batch123 batch of drugs were affected by during the production, transportation and storage process due to the temperature fluctuation of equipment A, how the abnormal propagation path was transmitted, and the impact weight of each node. This report will also mark the drug IDs that may need further inspection or recall to ensure that quality management personnel can take appropriate measures to deal with them in a timely manner.

[0029] Further, the S5 comprises the following steps: Based on the traceability analysis results, the corresponding quality-influencing parameters of the drugs affected by the anomaly are extracted, and the contribution of each parameter to the quality risk is calculated using the gradient boosting algorithm; Based on the training parameters selected by contribution, the XGBoost model is used to gradually train the corresponding training parameters of pharmaceutical products at production, transportation and storage nodes to generate quality risk scores for each node; Set risk thresholds through quality risk scoring and issue real-time risk warnings based on the risk thresholds.

[0030] In some embodiments, first, the system extracts the quality-influencing parameters corresponding to the drugs affected by the abnormal nodes based on the traceability analysis results. These parameters cover key variables in the production, transportation and storage links, such as production temperature, humidity, equipment working status, environmental control parameters (temperature, humidity, air pressure, etc.) during transportation, and storage conditions. After processing, these data will be used as input for subsequent analysis. For example, if a batch of drugs is affected by temperature fluctuations during the production process, the system will extract the temperature data of the equipment during the abnormal period, as well as other possibly related parameters in the same period, such as pressure, equipment speed, etc., and compare the quality changes under similar abnormal conditions through historical records to build a complete parameter set. Subsequently, the system uses the gradient boosting algorithm to calculate the extracted quality-influencing parameters to determine the contribution of each parameter to the overall quality risk. Gradient boosting is an iterative optimization algorithm that gradually reduces the error through the construction of multiple rounds of decision tree models. In this embodiment, the system first uses the quality-influencing parameters collected from each node as input features, and estimates the quality risk through the initial decision tree. Next, the system calculates the residual of the initial model (i.e., the difference between the actual risk and the predicted risk) and generates a new decision tree based on the residual to reduce the prediction error. Each iteration improves the accuracy of the prediction by adjusting the weights of each feature parameter. For example, if in a certain iteration, it is found that the fluctuation of the production temperature parameter contributes the most to the reduction of the residual, the weight of this parameter will be increased accordingly. Finally, through multiple iterations, the system can calculate the contribution of each parameter to the quality risk. After determining the contribution of each feature parameter, the system selects the training parameters based on these contributions and inputs them into the XGBoost model for step-by-step training. XGBoost is an efficient gradient boosting decision tree (GBDT) algorithm that can handle high-dimensional data and provide excellent model performance. During the training process, XGBoost gradually optimizes the risk score of each node by combining different quality-influencing parameters into multiple decision trees. Each node (production, transportation, storage) is trained based on its historical data and current input data to generate a specific risk score. XGBoost automatically adjusts the weights of features to ensure that important features (such as production temperature and transportation environment) are given higher priority, while minor features (such as irrelevant timestamps and invalid parameters) are downgraded. Ultimately, the system generates a quality risk score for each node, which reflects the potential threat of the node to drug quality. Finally, the system sets a reasonable risk threshold based on these risk scores. The setting of the risk threshold is based on the distribution of historical data and the actual frequency of quality problems. For each node, the system will determine the thresholds for different risk scores based on the actual occurrence of quality problems in historical data.When the risk score of a node exceeds the threshold, the system will trigger a real-time risk warning mechanism to notify relevant personnel to take corresponding measures. For example, if the risk score of a production node of a batch of drugs exceeds the set threshold (such as 0.85), the system will immediately issue an early warning to prompt the production supervisor to check the equipment status and adjust the production process to reduce the risk.

[0031] Furthermore, the formula of the XGBoost model is as follows: ; in, Score quality risks; is the output of the kth regression tree; is the weight of the corresponding quality influencing parameter; Quality influence parameter used for regression trees.

[0032] It should be noted that each regression tree can be understood as an independent model for predicting a specific risk. It will classify and weight these parameters through branch nodes according to quality-influencing parameters (such as production temperature, equipment pressure, transportation humidity, etc.). Ultimately, each leaf node of the regression tree represents a predicted value for quality risk. These predicted values ​​represent quality risks under a certain type of specific conditions. The output of the regression tree will not be used alone as the final result. The key to XGBoost is that it weights the output results of each regression tree through a combination of multiple regression trees to obtain an overall quality risk score. In other words, the prediction results of each regression tree are weighted according to their contribution to the overall risk. A regression tree with a larger weight represents a greater impact on the overall quality risk, while a regression tree with a smaller weight indicates that the tree has a lower contribution to risk prediction. In addition, the XGBoost model also has the ability to optimize step by step. During the training process of the model, it continuously adjusts the parameters of the regression tree so that the model can better fit the data and gradually reduce the prediction error. Each round of training will generate a new regression tree, which is specifically used to correct the prediction errors of all previous trees until the model reaches the best prediction accuracy. Ultimately, the weighted prediction results of all regression trees are combined into an overall quality risk score. This score indicates the quality problems that may occur in the production, transportation or storage of pharmaceutical products. The system sets a risk threshold based on historical data. When the score exceeds this threshold, the system will issue a quality risk warning to help relevant personnel take timely measures to ensure the safety and quality of the product.

[0033] The present invention also includes a pharmaceutical quality management system based on big data and artificial intelligence, including a big data acquisition module, a data cleaning module, an abnormality identification module, a traceability analysis module and a quality risk prediction module connected in sequence; A big data acquisition module is used to obtain quality-influencing parameters of pharmaceutical products from production to transportation through multimodal data acquisition equipment, wherein the quality-influencing parameters include production environment parameters, production equipment working parameters, pharmaceutical ingredient data, pharmaceutical batch information, and transportation and storage status; The data cleaning module is used to clean the quality-influencing parameters and perform data standardization, and perform dimensionality reduction processing on the standardized quality-influencing parameters through principal component analysis and autoencoder; The anomaly identification module is used to perform time series analysis on the quality-influencing parameters after dimensionality reduction based on the long short-term memory network and identify abnormal nodes in the production and transportation process; The traceability analysis module is used to establish a path tracing model for the entire life cycle of pharmaceutical products based on a graph neural network. The path tracing model learns the correlation between nodes, traces the source of quality problems based on abnormal nodes, and infers the affected drug batches and IDs to obtain traceability analysis results. The quality risk prediction module is used to predict the quality risk of real-time quality influencing parameters based on the results of traceability analysis through the XGBoost model.

[0034] The above implementation modes are merely descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering and technical personnel in the field shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A pharmaceutical quality management method based on big data and artificial intelligence, characterized in that: The following steps are involved: S1. Obtaining quality-influencing parameters of pharmaceutical products from production to transportation through multimodal data acquisition equipment, wherein the quality-influencing parameters include production environment parameters, production equipment working parameters, pharmaceutical ingredient data, pharmaceutical batch information, transportation and storage status, and pharmaceutical quality test results; S2. Clean the data of the quality-influencing parameters and perform data standardization, and perform dimensionality reduction processing on the standardized quality-influencing parameters through principal component analysis and autoencoder; S3. Perform time series analysis on the quality-influencing parameters after dimension reduction based on long short-term memory networks to identify abnormal nodes in the production and transportation processes; S4. Establish a path tracing model for the entire life cycle of pharmaceutical products based on graph neural networks. The path tracing model learns the correlation between nodes, traces the source of quality problems based on abnormal nodes, and calculates the affected drug batches and IDs to obtain traceability analysis results. S5. Based on the results of traceability analysis, the quality risk prediction of real-time quality influencing parameters is carried out through the XGBoost model.

2. A pharmaceutical quality management method based on big data and artificial intelligence according to claim 1, characterized in that: The production environment parameters include the temperature, humidity and air pressure of the production environment, the production equipment working parameters include the operating parameters of the equipment in each link of the pharmaceutical product production, the drug ingredient data include the drug ingredient raw materials and the corresponding batches, the drug batch information includes the production batch and production date of the drug, the transportation and storage status includes the transportation route point, transportation and storage temperature, transportation and storage humidity and transportation and storage air pressure, and the drug quality inspection results include whether there is any abnormality in the drug quality.

3. The pharmaceutical quality management method based on big data and artificial intelligence according to claim 1, characterized in that: The S2 comprises the following steps: Apply semantic segmentation model to classify quality-influencing parameters and assign type labels; Apply dynamic data cleaning rules to remove noise and outliers for quality-influencing parameters of different types of tags; Setting a hierarchical normalization function to perform nonlinear normalization on production environment parameters and transportation state parameters; The quality influencing parameters of production, transportation and storage dimensions are clustered by principal component analysis and K-means clustering algorithm to obtain the central features. The core feature data is obtained by performing nonlinear mapping and dimensionality reduction on the central features through the autoencoder.

4. A pharmaceutical quality management method based on big data and artificial intelligence according to claim 3, characterized in that: The S3 includes: A long short-term memory network model is constructed based on the core feature data. The core feature data generated by the real-time quality influencing parameters are compared with the core feature data generated by the historical quality influencing parameters based on the long short-term memory network model to calculate the probability of quality anomaly and generate abnormal nodes in the production, transportation or storage process.

5. A pharmaceutical quality management method based on big data and artificial intelligence according to claim 4, characterized in that: The path tracking model for the entire life cycle of pharmaceutical products is established based on the graph neural network according to the abnormal nodes, including the following steps: Modeling the links in the life cycle of a pharmaceutical product as nodes in a graph neural network model, wherein the links in the life cycle of a pharmaceutical product include production equipment, transportation tools, and storage facilities; Based on the quality impact parameters collected at each node, a data association matrix between nodes is generated; Based on the graph neural network, deep learning is performed on the data association matrix between nodes. By recursively processing the input and output data of the nodes, the data flow between each node is extracted, the influence weights between the nodes are calculated, the path weight matrix is ​​generated, and the path tracking model is constructed through the path weight matrix.

6. A pharmaceutical quality management method based on big data and artificial intelligence according to claim 5, characterized in that: The path tracing model is used to learn the correlation between nodes, trace the source of the quality problem, and calculate the affected drug batches and IDs to obtain the traceability analysis results, including the following steps: Based on the path weight matrix of the path tracing model, starting from the abnormal node, the influence weight of each node on its upstream node is recursively calculated, and the propagation path of the abnormal node is identified by comparing historical data with real-time data; Based on the abnormal propagation path and the drug batch information of the node, batch tracking is performed to calculate the drug batches and IDs affected by the abnormality, and the traceability analysis results are summarized and generated.

7. A pharmaceutical quality management method based on big data and artificial intelligence according to claim 6, characterized in that: The S5 comprises the following steps: Based on the traceability analysis results, the corresponding quality-influencing parameters of the drugs affected by the anomaly are extracted, and the contribution of each parameter to the quality risk is calculated using the gradient boosting algorithm; Based on the training parameters selected by contribution, the XGBoost model is used to gradually train the corresponding training parameters of pharmaceutical products at the production, transportation and storage nodes to generate the quality risk score of each node; Set risk thresholds through quality risk scoring and issue real-time risk warnings based on the risk thresholds.

8. The pharmaceutical quality management method based on big data and artificial intelligence according to claim 1, characterized in that: The formula of the XGBoost model is as follows: ； in, Score quality risks; is the output of the kth regression tree; is the weight of the corresponding quality influencing parameter; Quality influence parameter used for regression trees.

9. A pharmaceutical quality management system based on big data and artificial intelligence, applied to a pharmaceutical quality management method based on big data and artificial intelligence as claimed in any one of claims 1 to 8, characterized in that: It includes a big data acquisition module, a data cleaning module, an anomaly identification module, a traceability analysis module and a quality risk prediction module which are connected in sequence; A big data acquisition module is used to obtain quality-influencing parameters of pharmaceutical products from production to transportation through multimodal data acquisition equipment, wherein the quality-influencing parameters include production environment parameters, production equipment working parameters, pharmaceutical ingredient data, pharmaceutical batch information, and transportation and storage status; The data cleaning module is used to clean the quality-influencing parameters and perform data standardization, and perform dimensionality reduction processing on the standardized quality-influencing parameters through principal component analysis and autoencoder; The anomaly identification module is used to perform time series analysis on the quality-influencing parameters after dimensionality reduction based on the long short-term memory network and identify abnormal nodes in the production and transportation process; The traceability analysis module is used to establish a path tracing model for the entire life cycle of pharmaceutical products based on a graph neural network. The path tracing model learns the correlation between nodes, traces the source of quality problems based on abnormal nodes, and infers the affected drug batches and IDs to obtain traceability analysis results. The quality risk prediction module is used to predict the quality risk of real-time quality influencing parameters based on the results of traceability analysis through the XGBoost model.

Images