Traditional Chinese medicine preparation process knowledge representation learning research model embedded with time dimension

Abstract

The invention provides a traditional Chinese medicine preparation process knowledge representation learning research model embedded with a time dimension, and aims to solve the complex time sequence dependence problem and the multi-source heterogeneous data integration problem in the traditional Chinese medicine preparation process. According to the model, a time sequence knowledge graph with time information is constructed, and the time dependency relationship among process steps, operation parameters and equipment information is accurately modeled. According to the method, firstly, a CASREL model is adopted for entity relation joint extraction, the defects of a traditional method in processing triple overlapping are overcome, and the data extraction accuracy is improved; thirdly, time information is embedded into the extracted triple, and a tetrad (entity, relation, entity and time) is generated so as to enhance the time sequence reasoning capability; time sequence reasoning is carried out through a TTransE model, and the time change rule in the technological process is accurately captured. The method breaks through the limitation of lack of time dimension in traditional data analysis, effectively improves the optimization and intelligence level of the traditional Chinese medicine preparation process, is applied to the fields of process adjustment, production scheduling, quality control and the like, can provide more accurate decision support, and pushes the traditional Chinese medicine preparation industry to develop towards the intelligent and refined direction.

Description

Technical Field

[0001] The present invention relates to various process flows, equipment operation monitoring and production management systems in the traditional Chinese medicine pharmaceutical industry, especially to pharmaceutical process flows with complex time sequence dependencies. Background Art

[0002] The Chinese medicine pharmaceutical process covers the processing flow from Chinese medicinal materials to various finished medicines. In this process, in order to ensure the high quality of production, real-time monitoring equipment is widely deployed in various links of the production line. The entire production process involves many complex processing steps, and each process contains a variety of parameters that affect product quality. There are complex correlations between these parameters, which are intertwined and jointly affect the quality of the final product. On-site operators adjust these parameters according to established production specifications and process standards to solve quality problems and ensure the smooth operation of the production line.

[0003] However, at present, many traditional Chinese medicine manufacturing processes still rely too much on manual experience, the correlation between process flows is very complex, data often has interruptions, and management is also insufficient. Due to these problems, data islands often appear in production, making it difficult to integrate and utilize them efficiently. These problems are also closely related to factors such as the quality of medicinal materials, abnormalities in production equipment, temperature and humidity fluctuations in the operating environment, equipment wear and changes in its operating status, process changes, changes in the operating environment, material changes, and differences in operator experience. These factors may affect the process flow, and thus affect the quality and stability of the drug.

[0004] In order to deal with the above problems, traditional Chinese medicine pharmaceutical process data analysis methods mostly rely on manual experience to identify and deal with problems that arise in the production process. Since these methods often lack sufficient consideration of the time dimension in the process, they have defects such as low efficiency and poor accuracy when faced with complex time dependencies and multi-factor interactions. Moreover, traditional process data analysis methods are usually limited to means such as retrospective historical data and manual analysis, which makes it difficult to effectively extract and organize unstructured and semi-structured data generated in a large number of production processes, such as operating standards, batch records, process manuals, equipment logs and other information. Since this knowledge is difficult to systematize and standardize, it increases the complexity of staff in using and managing this information.

[0005] In response to these problems, the present invention proposes "a learning research model for traditional Chinese medicine pharmaceutical process knowledge representation embedded in the time dimension". This model can accurately model various parameters, operation steps and their sequential changes in the production process by effectively embedding time information. By constructing a knowledge representation with temporal dependencies, the model can not only deeply explore the time variation rules in the process, but also provide more efficient and accurate process parameter adjustment strategies and production process optimization solutions through precise modeling of the time dimension. In this way, not only can the dependence of traditional manual experience on process adjustment be effectively reduced, but also intelligent prediction of process flow changes can be achieved in a shorter time, helping the production line to achieve more efficient management and optimization. Summary of the invention

[0006] In view of the diversity and complexity of existing Chinese medicine pharmaceutical process data, there is currently a lack of public entity relationship triple extraction datasets, and traditional information extraction methods have limitations when dealing with multi-source heterogeneous data and triple overlaps. This paper proposes a Chinese medicine pharmaceutical process knowledge representation learning research model embedded with time dimension, aiming to enhance the temporal reasoning ability of the process by constructing a temporal knowledge graph with time information, thereby providing intelligent support for process optimization and adjustment.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] 1. Procurement and construction of data sets: Due to the complexity of Chinese medicine pharmaceutical process data, the public entity relationship triple extraction data set is still incomplete. To fill this gap, the present invention collects and integrates multi-source heterogeneous data from pharmaceutical process manuals, batch records, equipment operation logs, and Chinese medicine pharmaceutical literature, performs data preprocessing, and extracts core knowledge including process steps, operating parameters, equipment information, etc. By annotating these data, a high-quality data set suitable for entity relationship extraction is constructed. This data set not only covers the key information in the Chinese medicine pharmaceutical process, but also ensures the accuracy and comprehensiveness of the data through strict annotation specifications, providing a reliable basis for subsequent model training and experiments.

[0009] 2. Information extraction model selection: In the task of entity relationship extraction of traditional Chinese medicine pharmaceutical process data, traditional extraction methods have limited effect when dealing with multi-source heterogeneous data and triple overlapping problems. The present invention adopts the CASREL model for joint extraction of entity relationships. The CASREL model effectively reduces information loss by extracting entities and relationships at the same time, especially when dealing with triple overlapping relationships, which can improve the accuracy of extraction. The model can accurately extract core entities such as process steps, operating parameters, equipment and their relationships in traditional Chinese medicine pharmaceutical process data, and provide efficient and accurate data support for subsequent process analysis and optimization. The details are as follows:

[0010]

[0011] Among them, h is the hidden state of the model, w is the weight of the extracted entities and relations, and P(E,R) is the joint probability of entities and relations; CASREL's joint extraction method can effectively handle overlapping relationships in traditional Chinese medicine pharmaceutical data and improve extraction accuracy, thus providing a solid foundation for knowledge representation and temporal reasoning.

[0012] 3. Time information embedding and quadruple construction: In order to further enhance the temporal reasoning capability of the model, the present invention proposes to embed time information into entity relationship extraction and construct a temporal knowledge graph with time information. Specifically, based on the triples successfully extracted in the early stage, the present invention focuses on how to effectively embed the time dimension on the basis of the triples, thereby expanding them into quadruple. The core idea of ​​this method is to add time information based on the triples (entity, relationship, entity) successfully extracted in the early stage, and expand them into quadruple (entity, relationship, entity, time). The key to this step is to extract time information related to the process from the text and standardize it, thereby enhancing the temporal reasoning and semantic understanding capabilities of the model.

[0013] Quadruple=(E 1 ,R,E 2 ,T)

[0014] Among them, E1 and E2 are entities of the process steps, RRR is the relationship between them, and TTT is a time entity, indicating the time or time period when the process step occurs.

[0015] Furthermore, based on the CASREL model, a time information extraction module is added, which is responsible for extracting time information related to the process from the text. Through this module, the model can accurately identify time entities and extract their specific locations and semantic information in the text. Afterwards, the time information is combined with the triples to form a quadruple (entity, relationship, entity, time), which is used as the input of the TTransE model. Specifically, the time information extraction module relies on the pre-trained BERT model and is fine-tuned in combination with a specific time information annotation dataset. First, the text of the traditional Chinese medicine pharmaceutical process is standardized to ensure that the text can adapt to the input requirements of the BERT model. Then, the time entity is extracted and BERT is fine-tuned. The text input will first be converted into a token through the WordPiece Tokenizer. Usually in BERT, each word or subword will be mapped to a vector of fixed dimension. These vectors are represented as features in the input of BERT. For example, "October 10, 2023" may be split into ["2023", "year", "10", "month", "10", "day"], and then converted into a vector through BERT's WordPiece embedding. Then the sequence labeling task is used to fine-tune BERT. The input is a text sequence after word segmentation, and the label is the time entity category corresponding to each Token (such as "B-TIME" indicates the beginning of the time entity, "I-TIME" indicates the inside of the time entity, and the others are "O"). Next, the training data, for example, the label is: Token: "October 10, 2023"->Label: "B-TIME I-TIME I-TIME I-TIME". Use cross-entropy loss to train the model:

[0016]

[0017] Where yi is the true label (e.g. "B-TIME") and pi is the predicted probability output by BERT.

[0018] Furthermore, the BERT model outputs the label of each Token, which is further optimized through the CRF (Conditional Random Field) layer to identify the accurate boundaries of time entities. For example, the model labels "October 10, 2023" as "B-TIMEI-TIME I-TIME I-TIME", and can extract the time entity "October 10, 2023" from it. Next, the relative time (such as "5 days later") is standardized and converted into absolute time. This process depends on the current date and the prompt information in the text. For example, if the current date is "October 10, 2023", and the text mentions "5 days later", it is converted to "October 15, 2023" through simple date addition. Through time information extraction and standardization, the model can embed time information into triples and generate quadruplets with time information as input to the TTransE model.

[0019] Furthermore, the TTransE model can understand the semantics of temporal information and its sequential relationship, and analyze the impact of time on entities and relationships. In order to enhance the temporal reasoning ability of the model, this paper introduces the TTransE model, which can perform temporal reasoning based on the quadruple (entity, relationship, entity, time). Its goal is to learn the association between entities, relationships, and temporal information by minimizing the energy function of the model:

[0020]

[0021] Among them, h, r, t represent the head entity, relationship and tail entity respectively, T is the time information, γ(h, r, t, T) is the penalty term for negative samples, and the L norm is used to measure the distance between entities and time information. After embedding time information, the model can better adapt to complex Chinese medicine pharmaceutical process data, help optimize process parameters, adjust production scheduling, and enhance the guarantee of knowledge fusion and information consistency.

[0022] The beneficial effects of the present invention are:

[0023] The proposed "a time-embedded TCM pharmaceutical process knowledge representation learning research model" significantly improves the temporal reasoning and knowledge representation capabilities of the process by effectively embedding time information in the analysis of traditional TCM pharmaceutical process data. The model integrates key information from multi-source heterogeneous data (such as process manuals, batch records, equipment logs, etc.) and uses the CASREL model to jointly extract entities and relationships, thereby solving the limitations of traditional methods such as data islands and overlapping triples, thereby improving the accuracy and comprehensiveness of data extraction.

[0024] In addition, the present invention combines BERT and CRF technologies to accurately extract and standardize time information from the TCM pharmaceutical process text, further enriching the process knowledge representation, so that the time dimension can be accurately integrated into the triple and converted into a quadruple (entity, relationship, entity, time). This innovative time series knowledge graph provides the model with a deeper time series understanding capability, which can capture the laws of changes in process parameters, equipment status and operation steps over time.

[0025] By introducing the TTransE model, this method can not only conduct deep learning of time information in the process, but also perform accurate time series reasoning based on the passage of time to optimize process parameter adjustment, production scheduling and process control. This greatly improves the automation and intelligence level of the production process, reduces excessive reliance on manual experience, and can accurately predict process changes in a short period of time, improve production efficiency and ensure the stability of product quality.

[0026] In general, the innovative application of this invention in the field of traditional Chinese medicine pharmaceutical manufacturing not only provides an efficient process optimization and adjustment plan, but also enhances the integration and application capabilities of process knowledge through a systematic and standardized knowledge representation method, providing solid technical support for the intelligent and digital transformation of the traditional Chinese medicine pharmaceutical industry, and has significant theoretical value and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a data flow and data source structure diagram of the Chinese medicine pharmaceutical process of the present invention;

[0028] Figure 2 It is a flow chart of entity relationship joint extraction of CASREL model of the present invention;

[0029] Figure 3 It is a schematic diagram of time information embedding and quaternary construction of the present invention;

[0030] Figure 4 It is a timing reasoning and process optimization process diagram of the TTransE model of the present invention;

[0031] Figure 5 It is a system architecture diagram of the present invention. DETAILED DESCRIPTION

[0032] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0033] It should be noted that, unless otherwise specified, all technical and scientific terms used in this application have the same meanings as commonly understood by ordinary technicians in the technical field to which this application belongs.

[0034] In the present invention, unless otherwise specified, the directions used, such as "up" and "down", usually refer to the directions shown in the drawings, or to the vertical, perpendicular or gravity directions; similarly, for ease of understanding and description, "left" and "right" usually refer to the left and right shown in the drawings; "inside" and "outside" refer to the inside and outside relative to the outline of each component itself, but the above-mentioned directions are not used to limit the present invention.

[0035] Reference Figure 1 ,This paper proposes a learning model of Chinese medicine pharmaceutical ,processing module, knowledge representation module, temporal reasoning module and process optimization suggestion module. The data collection module is responsible ,for collecting Chinese medicine pharmaceutical process data from multiple data sources (such as process manuals, equipment operation logs, ,batch records, etc.). Figure 1 The flow of these data sources is shown, with special emphasis on the process of data preprocessing. First, all collected data needs to be processed through steps such as data cleaning, denoising, and standardization. Then, data from different sources are integrated into a unified input data set as the basis for subsequent processing. This data includes but is not limited to raw material information, operation steps, equipment status, time stamps, etc.

[0036] Reference Figure 2 The system uses the CASREL model to perform knowledge representation and relationship extraction on the Chinese medicine pharmaceutical process data. The CASREL model can not only extract core entities such as process steps, operating parameters, and equipment, but also extract the relationships between them. This step converts the data from the original text form into a structured triple representation, which facilitates subsequent knowledge reasoning and analysis. Figure 2 The workflow of the CASREL model is clearly demonstrated, explaining how to implement the joint task of entity extraction and relationship extraction in the same model. Through this model, various key information and relationships in the traditional Chinese medicine manufacturing process can be effectively captured, providing data support for process optimization.

[0037] Reference Figure 3 ,The system further expands the traditional triple to quad ,tuple to capture the time dimension in process data by embedding ,time information. Figure 3 It shows in detail how to embed time information with entities and relations, and expand triples into quadruples (entity, relation, entity, time). In this process, the model associates process operations at different time points with time series by extracting and standardizing timestamps, thereby providing more accurate input data for time series reasoning and process optimization. This design can effectively ensure the importance of time factors in the process and provide a basis for subsequent reasoning and optimization processes.

[0038] Reference Figure 4 ,TTransE model is used to perform temporal reasoning on quadruple data with time information. Figure 4 It shows how the TTransE model learns the associations between entities, relationships, and time based on minimizing the energy function. Specifically, the model analyzes data such as process parameters and operation steps at different time points to infer the optimal process parameters at different times. This reasoning process can not only optimize the current process operation, but also predict future process trends, thereby providing timely and accurate adjustment suggestions for the production line to optimize production efficiency and drug quality.

[0039] Reference Figure 5 The overall architecture design of the system is as follows Figure 5 As shown in the figure, it includes data input layer, data processing layer, knowledge extraction module, time series reasoning module and process optimization decision output layer. The modules are interconnected through data streams to form an efficient data processing and analysis system. The data input layer is responsible for receiving data from different data sources and preprocessing it through the data processing layer. The knowledge extraction module implements data structuring and knowledge representation through the CASREL model. The time series reasoning module performs reasoning based on the TTransE model, and finally feeds back the optimization results to the operators through the process optimization decision output layer to help them make more scientific process adjustment decisions.

[0040] The embodiments of the present invention provide a method for optimizing the pharmaceutical process of traditional Chinese medicine that integrates deep learning, temporal reasoning and knowledge representation technology. Through accurate time information embedding and temporal reasoning models, automatic optimization of process parameters can be achieved, and process adjustment suggestions with real-time feedback can be generated, significantly improving the automation level and drug quality in the production process.

[0041] It should be emphasized that although the present invention describes the technical solution in detail through multiple embodiments, those skilled in the art should understand that the technical solution in the embodiments can be modified or replaced by equivalents without departing from the spirit of the present invention. These modifications or replacements will not change the core technical content and innovation of the present invention. Therefore, the protection scope of the present invention shall be subject to the claims.

Claims

1. A research model for learning knowledge representation of traditional Chinese medicine pharmaceutical process embedded in the time dimension, characterized by: The model accurately models the time dependencies among process steps, operating parameters and equipment information in the Chinese medicine manufacturing process by constructing a temporal knowledge graph with time information, and provides process parameter adjustment and production process optimization solutions based on temporal reasoning.

2. According to a time-dimensional embedded traditional Chinese medicine pharmaceutical process knowledge representation learning research model according to claim 1, it is characterized in that: The model includes the following steps: (1) Using multi-source heterogeneous data sets, we collected and integrated data from traditional Chinese medicine manufacturing process manuals, batch records, equipment operation logs, and traditional Chinese medicine manufacturing literature, and performed data preprocessing to extract core knowledge such as process steps, operating parameters, and equipment information; (2) The CASREL model is used to perform entity-relationship joint extraction to extract the entities and their relationships of process steps, operating parameters, and equipment information in the traditional Chinese medicine pharmaceutical process.

3. According to a time-dimensional embedded traditional Chinese medicine pharmaceutical process knowledge representation learning research model according to claim 2, it is characterized in that: The model embeds time information in the triples to generate quadruples (entity, relationship, entity, time) to enhance the temporal reasoning capability.

4. According to a time-dimensional embedded traditional Chinese medicine pharmaceutical process knowledge representation learning research model according to claim 3, it is characterized in that: The time information is extracted from the traditional Chinese medicine pharmaceutical process text through the time entity extraction module, and the BERT model is fine-tuned in combination with the time annotation dataset in the text to identify and standardize the time entities and generate accurate quadruplets.

5. According to claim 4, a time-dimensional embedded traditional Chinese medicine pharmaceutical process knowledge representation learning research model is characterized in that: The standardization of the time entity is achieved by processing relative time (such as "5 days later") and converting it into absolute time, thereby ensuring the consistency and accuracy of time information.

6. According to a time-dimensional embedded traditional Chinese medicine pharmaceutical process knowledge representation learning research model according to claim 1, it is characterized in that: The model performs timing reasoning through the TTransE model to understand the semantics of time information and its impact on process steps and parameter adjustments.

7. According to a time-dimensional embedded traditional Chinese medicine pharmaceutical process knowledge representation learning research model according to claim 1, it is characterized in that: The model can be applied to the optimization of traditional Chinese medicine pharmaceutical process, production scheduling, quality control and equipment maintenance, and provide intelligent support.

Images