CN118152481B

CN118152481B - Drug information storage method based on distributed edge calculation and multi-mode data

Info

Publication number: CN118152481B
Application number: CN202410572705.4A
Authority: CN
Inventors: 刘万里; 陈瑞香; 段学涛
Original assignee: Tianjin Minxiang Biomedical Co ltd
Current assignee: Tianjin Minxiang Biomedical Co ltd
Priority date: 2024-05-10
Filing date: 2024-05-10
Publication date: 2024-08-09
Anticipated expiration: 2044-05-10
Also published as: CN118152481A

Abstract

The invention relates to the technical field of information processing, in particular to a medicine information storage method based on distributed edge calculation and multi-mode data, which comprises the following steps: step S1: selecting an embedded system suitable for edge computing as edge equipment, and customizing an operating system and a software stack of the equipment to meet the requirements of drug information acquisition, storage and processing; step S2: collecting multi-modal data of a drug, wherein the multi-modal data comprises text data, image data and video data; step S3: and establishing a multi-mode data fusion and processing module on the edge equipment, and processing and analyzing the acquired multi-mode data in real time by utilizing a lightweight deep learning model. According to the invention, through a collaboration and optimization mechanism between the edge devices and intelligent distribution and scheduling of the edge computing tasks based on deep reinforcement learning, interconnection and intercommunication of data and resource sharing are promoted, and efficient integration and unified management of drug information are realized.

Description

Drug information storage method based on distributed edge calculation and multi-mode data

Technical Field

The invention relates to the technical field of information processing, in particular to a medicine information storage method based on distributed edge calculation and multi-mode data.

Background

At present, the existing storage method of drug information mainly comprises the following steps:

And (3) storing a database: the drug information may be stored in various databases, such as a relational database (e.g., mySQL, oracle, etc.) or a non-relational database (e.g., mongoDB, redis, etc.). The databases can establish corresponding table structures or document structures according to the attributes (such as names, components, doses, indications, adverse reactions and the like) of the medicaments, and the data can be conveniently queried, retrieved and managed.

And (3) storing files: the medication information may also be stored in a local file system or cloud storage service in the form of files, with common formats including text files (e.g., CSV, JSON, XML, etc.) and binary files. These files may be organized in a structure to facilitate program reading and parsing of information therein.

Knowledge graph: knowledge graph is a graphical data structure for representing and storing knowledge, and drug information can be stored and managed in the form of graph. Various attributes and relationships of the medicine can be expressed as nodes and edges in the graph, so that semantic query and reasoning are facilitated.

Although the existing drug information storage method has certain advantages, the existing drug information storage method also has some defects, including the following points:

1. data islanding problem: drug information is often scattered in different databases, files or systems, with the situation of data islands. This results in fragmentation and inconsistency of the data, increasing the difficulty of data management and integration.

2. Data normalization: the standardization of drug information is not high, and different data sources and systems may adopt different data formats and naming standards, resulting in inconsistent data and difficult integration. The lack of a unified data standard makes exchange and sharing of data difficult.

3. Query efficiency: for large-scale pharmaceutical information data, traditional database query and retrieval methods are inefficient, especially in the case of complex queries and across data sources. In addition, conventional data storage methods cannot meet the storage and processing requirements of large-scale data.

There is therefore a need for a method of drug information storage based on distributed edge computing and multimodal data that addresses the above-described problems.

Disclosure of Invention

The invention provides a medicine information storage method based on distributed edge calculation and multi-mode data, which promotes the interconnection and the resource sharing of data and realizes the efficient integration and the unified management of medicine information through a collaboration and optimization mechanism between edge devices and the intelligent distribution and the scheduling of edge calculation tasks based on deep reinforcement learning.

The technical scheme adopted by the invention for solving the technical problems is as follows: a drug information storage method for calculating and multi-modal data at a distributed edge comprises the following steps:

Step S1: selecting an embedded system suitable for edge computing as edge equipment, and customizing an operating system and a software stack of the equipment to meet the requirements of drug information acquisition, storage and processing;

step S2: collecting multi-modal data of a drug, wherein the multi-modal data comprises text data, image data and video data;

Step S3: establishing a multi-mode data fusion and processing module on the edge equipment, and processing and analyzing the acquired multi-mode data in real time by utilizing a lightweight deep learning model;

step S4: correlating text data, image data and video data through a fusion model of a deep neural network, and extracting key features of drug information;

Step S5: based on storage resources and computing power of the edge equipment, designing a self-adaptive data storage and management strategy;

step S6: adopting a storage optimization algorithm based on deep reinforcement learning, and dynamically adjusting a data storage and index strategy according to the importance and timeliness of the drug information;

Step S7: establishing a coordination and optimization mechanism among distributed edge computing nodes in the edge equipment, and realizing dynamic scheduling and coordination computing of edge computing resources;

step S8: the intelligent allocation and the scheduling of the edge calculation tasks are realized through an edge calculation collaborative optimization algorithm based on deep reinforcement learning;

Step S9: storing the drug information on a plurality of nodes in a distributed network to realize decentralization storage;

step S10: and defining an access control strategy on the distributed network through an intelligent contract, and encrypting the medicine information by adopting a homomorphic encryption method.

Further, in the step S1, selecting a lightweight, low-power embedded system suitable for edge computing as an edge device includes:

Step S1-1: demand analysis:

Determining specific scenes and requirements of drug information acquisition, including acquisition environment, data type, real-time requirements and safety requirements;

the functional requirements of the edge equipment are defined, including data acquisition, local storage and real-time processing;

Step S1-2: edge device selection:

according to the requirement analysis result of the step S1-1, selecting edge equipment, wherein the edge equipment adopts any one of Raspberry Pi or NVIDIA Jetson Nano embedded system based on ARM architecture.

Further, the operating system and the software stack of the customizing device in step S1 include:

Step S1-3: and (3) customizing design:

And (3) carrying out customized design on the edge equipment selected in the step S1-2, wherein the customized design comprises an operating system and a software stack of the customized equipment so as to meet the requirements of medicine information acquisition, storage and processing, and the customized content comprises the following steps: selecting an operating system, configuring lightweight data acquisition and storage services and integrating an edge computing framework;

step S1-4: performance evaluation and testing:

After the customized design is completed, performance evaluation and testing are carried out, stability and performance of the edge equipment under different environments are tested, the edge equipment can meet the requirements of drug information management, and the testing contents comprise: the accuracy and the instantaneity of data acquisition, the stability and the safety of data storage and the efficiency and the accuracy of data processing are all achieved, and performance evaluation and test results are obtained;

step S1-5: optimizing and adjusting:

And (3) optimizing and adjusting the edge equipment according to the performance evaluation and test results obtained in the step (S1-4), wherein the optimization and adjustment comprise system configuration adjustment, software algorithm optimization and data transmission and storage mechanism improvement.

Further, the establishing the multi-mode data fusion and processing module on the edge device in the step S3 includes:

Step S3-1: determining a resource limit of the edge device:

Determining hardware resource limitation of the edge equipment, including processor performance, memory capacity, storage space and energy consumption, and knowing network connection condition of the edge equipment to determine bandwidth and delay of data transmission;

Step S3-2: analyzing the multimodal data types:

Analyzing multi-mode data types to be processed on the edge equipment, and selecting a processing method and an algorithm according to the characteristics and the processing requirements of each data type;

step S3-3: designing a data input interface:

a unified data input interface is designed and used for receiving multi-mode data from different data sources, and determining the format and protocol of data input, and the mode and rate of data transmission;

Step S3-4: data preprocessing:

Preprocessing input multi-mode data to reduce noise and improve data quality;

Step S3-5: multimodal data fusion:

and fusing the preprocessed multi-modal data, constructing a unified data representation form, and then selecting a lightweight deep learning model to process and analyze the acquired multi-modal data in real time.

Further, the selecting the lightweight deep learning model to process and analyze the collected multi-modal data in real time includes:

Step S3-6: preprocessing the acquired multi-mode data, including size adjustment of image data, word segmentation and vectorization of text data and frame extraction of video data, so as to ensure that the format and representation of the data meet the input requirements of a model;

Step S3-7: selecting a lightweight deep learning model:

Selecting a proper lightweight deep learning model according to the resource limitation and real-time processing requirement of the edge equipment, wherein the lightweight deep learning model adopts more than three learning models of MobileNet, squeezeNet, LSTM, GRU, distilBERT, temporal Convolutional Networks and TCN;

step S3-8: model deployment and optimization:

After a lightweight deep learning model is selected, deploying the lightweight deep learning model on edge equipment for real-time processing, and reducing the size and the computational complexity of the model by using model quantization, pruning and model distillation so as to improve the running efficiency of the model on the edge equipment;

step S3-9: real-time processing and analysis:

The collected data are processed step by adopting a stream processing method, and the collected multi-mode data are input into a model to be inferred, so that a corresponding output result is obtained.

Further, the key features of the drug information in step S4 include a drug identification feature, a drug text information feature, a drug usage scenario feature and a drug correlation feature.

Further, in the step S5, based on the storage resources and computing power of the edge device, designing the adaptive data storage and management policy includes:

step S5-1: firstly, comprehensively identifying hardware configuration of edge equipment, wherein the hardware configuration comprises, but is not limited to, CPU type, core number, main frequency, memory capacity, hard disk type, capacity and network bandwidth parameters;

step S5-2: continuously monitoring real-time load conditions of the edge equipment, including CPU (Central processing Unit) utilization rate, memory occupancy rate, disk I/O (input/output) and network flow, so as to accurately master actual use conditions and residual capacity of equipment resources;

step S5-3: counting the total amount, growth rate and proportion of different types of data of the drug information so as to predict future storage requirements;

step S5-4: through log recording and data analysis, the access frequency, the query mode and the read-write proportion characteristic of the drug information data are known so as to optimize the data layout and the index strategy;

Step S5-5: identifying the validity period, update period and importance grade attribute of the drug information data, and providing basis for formulating data aging, archiving and deleting strategies;

Step S5-6: distributing data among the cache, the fast storage and the mass storage according to the data access speed requirement and the storage cost requirement;

step S5-7: according to the data scale and the storage capacity of the edge equipment, designing a data slicing and partitioning scheme;

step S5-8: and acquiring a data access mode analysis result, and dynamically adjusting the layout of the data in the storage medium by combining the real-time resource monitoring data of the edge equipment.

Further, in the step S6, a storage optimization algorithm based on deep reinforcement learning is adopted, and according to the importance and timeliness of the drug information, dynamically adjusting the data storage and index policy includes:

Step S6-1: the method comprises the steps of defining storage optimization targets, namely minimizing storage cost, maximizing access performance, guaranteeing data reliability, setting a composite objective function, and comprehensively considering a plurality of targets;

Step S6-2: defining state variables of storage optimization problems, including importance and timeliness of current drug information data, use condition of storage resources, data access mode and system load;

Step S6-3: selecting a state transition model to describe a law of state change in the storage environment;

Step S6-4: determining an action space for storing the optimization problem;

step S6-5: constructing a deep reinforcement learning model to capture time series characteristics of data importance and timeliness; step S6-6: a simulator of the storage environment is constructed using the historical data or the synthetic data to facilitate generating interaction samples during model training.

Further, the establishing a collaboration and optimization mechanism between distributed edge computing nodes in the edge device in step S7 includes:

step S7-1: defining edge computing node roles and communication protocols;

step S7-2: constructing a distributed collaborative framework;

step S7-3: and realizing cooperative calculation.

Further, the edge computing collaborative optimization algorithm based on the deep reinforcement learning in the step S8 includes:

Step S8-1: the type and the characteristics of the edge computing task are defined, and the resource requirements, the task priority and the time constraint are computed;

step S8-2: constructing a state model for reflecting the whole edge computing system;

step S8-3: designing a deep reinforcement learning model framework;

step S8-3: training a deep reinforcement learning model;

step S8-4: deployment and real-time decision making.

The invention has the advantages that:

1. According to the invention, the distributed edge computing network is constructed, the drug information is stored on a plurality of nodes in the network, the nodes in the distributed edge computing network follow a unified data model, interface standards and protocols, the standardization and consistency of data in form are ensured to be logically coherent and interoperable, the problem of logic isolation of data islands is eliminated, and the interconnection and intercommunication of data and resource sharing are promoted and the efficient integration and unified management of the drug information are realized through a collaboration and optimization mechanism between edge devices and intelligent distribution and scheduling of edge computing tasks based on deep reinforcement learning.

2. The invention adopts a storage optimization algorithm based on deep reinforcement learning, and can dynamically adjust the data storage and index strategy according to the importance and timeliness of the drug information, thereby ensuring that the key data is protected preferentially. Meanwhile, an access control strategy is defined on a distributed network by intelligent combination, and the homomorphic encryption method is combined to encrypt the drug information, so that the safety of data is enhanced, the data is effectively prevented from being leaked, tampered and lost, and a higher-level security guarantee is provided especially in the distributed storage and cloud storage environments.

3. The invention ensures that the collection, storage and processing of the drug information follow the unified standard and specification by customizing the operating system and the software stack of the edge equipment. The multi-mode data fusion and processing module utilizes a lightweight deep learning model to perform standardized processing and association analysis on drug data of different sources and different modes, extracts key features, improves the consistency and standardization degree of the data, and is beneficial to data exchange and sharing.

4. The fusion model of the deep neural network can effectively correlate text data, image data and video data, extracts key characteristics of drug information, and lays a foundation for efficient query and retrieval. And the self-adaptive data storage and management strategy combines the storage resources and calculation power of the edge equipment, so that the storage and processing efficiency of large-scale medicine information data is improved. Particularly, for the complex inquiry and the case of crossing data sources, the scheme can obviously improve the inquiry response speed and meet the storage and processing requirements of large-scale data.

In summary, the invention successfully solves the key problems in the existing drug information storage method through the technical means of distributed edge calculation, multi-mode data fusion, deep learning model application, intelligent storage optimization, data encryption, access control and the like, realizes the efficient integration, safe storage, standardized processing and quick query of the drug information, and greatly improves the overall performance and value of the drug information management system.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are needed in the description of the embodiments or the prior art will be briefly described, and it is obvious that the drawings in the description below are some embodiments of the present invention, and other drawings can be obtained according to the drawings without inventive effort for a person skilled in the art.

Fig. 1 is a flowchart of a drug information storage method based on distributed edge computing and multi-modal data provided by the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made apparent and fully in view of the accompanying drawings, in which some, but not all embodiments of the invention are shown. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

Example 1: fig. 1 is a flowchart of a method for storing drug information based on distributed edge computing and multi-mode data according to the present invention, where the method for storing drug information based on distributed edge computing and multi-mode data shown in fig. 1 includes the following steps:

the medicine information storage method based on the distributed edge calculation and the multi-mode data is characterized by comprising the following steps:

Step S1: considering that the drug information can be collected and processed in various environments, a lightweight and low-power-consumption embedded system suitable for edge computing is selected as an edge device, such as an ARM architecture-based embedded system. Meanwhile, the operating system and the software stack of the equipment are customized to meet the requirements of drug information acquisition, storage and processing;

An Embedded System (Embedded System) refers to a special purpose computer System designed for a specific application or task, and comprises two major parts, i.e., hardware (e.g., a processor, a memory, a peripheral interface, etc.) and software (e.g., an operating System, an application program, etc.). Embedded systems are typically integrated into larger physical devices that provide control, monitoring, data processing, etc. functions for the device.

Edge devices (EDGE DEVICE) refer to computing devices that perform data processing, analysis, and storage tasks at the edge of the network (near the data generation source or user side). The cloud computing center and the terminal user or the Internet of things equipment are positioned between the cloud computing center and the terminal user or the Internet of things equipment, and are used for carrying out tasks such as data preprocessing, real-time response, privacy protection and the like, so that delay is reduced, cloud burden is lightened, and data safety is guaranteed

Because the embedded system has a complete hardware platform and a customized software environment, specific computing tasks can be independently completed, and therefore, the embedded system completely accords with the definition of edge equipment in function. In many edge computing scenarios, it is various embedded systems that are used, such as intelligent gateways, industrial controllers, internet of things sensor nodes, smart home devices, and the like. The embedded systems integrate necessary computing, communication and storage capacity at a hardware level, and an operating system, middleware and application programs which are suitable for the edge computing requirements are installed at a software level, so that the embedded systems together form entity equipment for executing the edge computing tasks.

In the step S1, selecting a lightweight and low-power embedded system suitable for edge computing as an edge device includes:

Step S1-1: demand analysis:

Determining specific scenes and requirements of drug information acquisition, including acquisition environment, data type, real-time requirements and safety requirements; the functional requirements of the edge equipment are defined, including data acquisition, local storage and real-time processing;

Step S1-2: edge device selection:

According to the requirement analysis result of the step S1-1, selecting edge equipment, wherein the edge equipment adopts any one of Raspberry Pi or NVIDIA Jetson Nano embedded system based on ARM architecture, and the equipment has the characteristics of light weight and low power consumption so as to ensure stable operation in various environments.

The operating system and software stack of the customizing device in step S1 include:

Step S1-3: and (3) customizing design: and (3) carrying out customized design on the edge equipment selected in the step S1-2, wherein the customized design comprises an operating system and a software stack of the customized equipment so as to meet the requirements of medicine information acquisition, storage and processing, and the customized content comprises the following steps: selecting an operating system such as a Linux system; configuring lightweight data acquisition and storage services and integrating an edge computing framework; such as TensorFlow Lite, openVINO, etc.

Step S1-4: performance evaluation and testing:

step S1-5: optimizing and adjusting:

Optimizing and adjusting the edge equipment according to the performance evaluation and test results obtained in the step S1-4, wherein the optimization and adjustment comprises system configuration adjustment, software algorithm optimization and data transmission and storage mechanism improvement; through continuous optimization, the edge equipment can be ensured to stably and efficiently run in various environments, and the actual requirements of drug information management are met. Through the steps, proper edge equipment can be effectively selected and customized, the medicine information can be stably and efficiently collected, stored and processed in various environments, the purpose of processing the medicine information is to extract useful features in the medicine information so as to better store, analyze and apply the medicine information subsequently, and particularly, the medicine information possibly contains a large amount of data, but not all the data are necessary for the subsequent analysis and application, so that the medicine information needs to be subjected to feature extraction, and key features in the medicine information, such as fingerprints of chemical structures, descriptors of medicine activity and the like, are extracted, so that the data dimension is reduced and the useful information is reserved. The drug information may be in different data types and formats, requiring unified data conversion and encoding for subsequent storage and processing. This may involve converting the textual description into a vector representation, converting the chemical structure into a molecular fingerprint, and so on; sometimes, the drug information may need to be fused or enhanced with other related information to enrich the content of the data and improve the quality of the data. For example, drug information is fused with clinical trial data, drug interaction information, etc., to increase the value and range of applications of the data. Therefore, the processing of the drug information is to convert the raw data into a form more suitable for storage, analysis and application to meet the requirements of drug information collection, storage and processing.

Step S2: collecting multi-modal data of a drug, wherein the multi-modal data comprises text data, image data and video data; specifically, the method comprises the following steps:

Step S2-1: determining data sources and collection targets, and sorting and listing various sources that may contain drug multimodal data, such as official databases and literature: drug registration information issued by drug regulatory authorities, pharmacopoeias, clinical trial reports, academic journal articles, and the like. Such as product specifications disclosed by pharmaceutical enterprises and medical institutions, experimental records and image data in the process of drug development; such as drug evaluation, drug use hearts and drug pictures uploaded by users disclosed by an Internet disclosure platform.

Step S2-3: for each data source, specific multi-modal data types to be collected are defined, for example, text description and medicine pictures are obtained from medicine specifications; extracting curative effect evaluation text from pharmacopoeia; and capturing user evaluation text, uploaded medicine use photos and videos from an Internet platform.

Step S2-4: designing and implementing a data acquisition strategy, and acquiring image, text and video data:

The text data acquisition comprises: API interface call: for data sources (such as official databases, part of professional websites) that provide a public API interface, the authoring program obtains text data in bulk through the API interface.

Webpage crawler: and developing customized webpage crawlers aiming at data sources without API interfaces or access limitation, following the website robots. Manual entry or export: for data (such as internal systems and electronic medical records) which can be accessed only through a man-machine interaction interface, a worker can manually input or export the data into a processable text file. The image data acquisition includes: API interface call: if the data source provides the API interface of the image resource, the interface is directly called to download the picture. Webpage element analysis: and (5) positioning the picture URL for the image resource on the webpage through HTML analysis or a CSS selector, and downloading and storing. Screen shots or recordings: for the protected or dynamically displayed image, the screen capturing or recording operation is carried out under the condition of adhering to the using terms, and then the image processing is carried out to extract the required information. The video data acquisition includes: API interface call: and if the data source provides an API interface of the video resource, the calling interface acquires the video file or the streaming media address. Webpage element analysis: and analyzing the webpage video elements, extracting video URLs, and downloading video files through an HTTP/HTTPS protocol. And carrying out association integration on the multi-modal data according to the drug identification (such as drug name, batch number, unique ID and the like) to form a structured drug multi-modal data set.

In summary, step S2 systematically collects multi-modal data (text, image, video) of the drug by determining data sources and collection targets, designing and implementing data collection strategies, data preprocessing and integration, and other logic steps, thereby laying a foundation for subsequent edge calculation processing, feature extraction, information fusion, and other tasks.

Step S3: establishing a multi-mode data fusion and processing module on the edge equipment, and processing and analyzing the acquired multi-mode data in real time by utilizing a lightweight deep learning model; specifically, the method comprises the following steps:

Step S3-1: determining a resource limit of the edge device:

Step S3-2: analyzing the multimodal data types:

Analyzing multi-modal data types (text data, image data and video data) to be processed on the edge device, and selecting a processing method and algorithm according to the characteristics and processing requirements of each data type;

Different processing methods and algorithms can be selected to achieve efficient processing for the processing requirements and characteristics of text data, image data and video data on the edge device. Text data is typically composed of words, phrases or sentences, with a high degree of structuring, but there may also be some degree of noise and variation, and in particular the following models may be used to analyze the text data: bag of Words model (Bag-of-Words, boW) and TF-IDF: for text representation and feature extraction. Word embedding model (e.g., word2Vec, gloVe, BERT): for learning semantic representations of text. Deep learning models (e.g., convolutional neural network, cyclic neural network, transducer, etc.): end-to-end learning for text tasks.

Wherein the image data consists of pixels, the image data processing involves tasks such as image classification, object detection, image segmentation, image generation, etc., and the following models can be used to analyze the image data: convolutional Neural Network (CNN): for image feature extraction and representation learning. Target detection algorithms (e.g., YOLO, fast R-CNN, SSD, etc.): for detecting objects and objects in images. Image segmentation algorithms (e.g., mask R-CNN, U-Net, FCN, etc.): for segmenting the image into semantic regions.

The video data is composed of continuous image frames, has time sequence and continuity, and has the characteristic of image data. Video data processing involves tasks such as video classification, motion recognition, object tracking, video summarization, etc. The following models and algorithms may be used in particular to analyze video data: 3D convolutional neural network (3D CNN): image feature extraction and representation learning taking into account timing information; video summarization algorithms (e.g., key frame extraction, video summary generation, etc.): for extracting key information from the video.

Step S3-3: designing a data input interface:

A unified data input interface is designed for receiving multi-modal data from different data sources and determining the format and protocol of the data input, as well as the manner and rate of data transmission.

Step S3-4: data preprocessing:

Preprocessing input multi-mode data to reduce noise and improve data quality; the image is subjected to the processes of size adjustment, color space conversion, enhancement, denoising and the like. The video is subjected to segmentation, downsampling, frame rate control, key frame extraction and other processes. And performing word segmentation, word vectorization, word deactivation, standardization and other processing on the text.

Step S3-5: multimodal data fusion:

The preprocessed multi-mode data are fused to construct a unified data representation form, and data information of different modes can be fused together by adopting strategies such as series connection, parallel connection or attention mechanism. And then, a lightweight deep learning model is selected to process and analyze the acquired multi-mode data in real time.

Through the steps, the multi-mode data fusion and processing module can be established on the edge equipment, so that the real-time processing and analysis of the multi-mode data are realized, and support and foundation are provided for edge computing application.

The step S3 of selecting the lightweight deep learning model to process and analyze the acquired multi-mode data in real time comprises the following steps:

Step S3-7: selecting a lightweight deep learning model:

And selecting a proper lightweight deep learning model according to the resource limitation and the real-time processing requirement of the edge equipment, wherein the lightweight deep learning model adopts more than three learning models of MobileNet, squeezeNet, LSTM, GRU, distilBERT, temporal Convolutional Networks and TCN, and the models have smaller model parameter quantity and calculation complexity and are suitable for real-time processing in the environment with limited resources. Specifically, mobileNet, squeezeNet and other lightweight convolutional neural network models can be used for image data processing; for text data, a lightweight recurrent neural network (e.g., LSTM, GRU) or a model based on a transducer architecture (e.g., distilBERT) may be selected; for video data, a lightweight 3D CNN model or a temporal convolution network (Temporal Convolutional Networks, TCN) may be selected.

Step S3-8: model deployment and optimization:

step S3-9: real-time processing and analysis:

The acquired data is processed step by adopting a stream processing method, the acquired multi-mode data is input into a model to be inferred, a corresponding output result is obtained, and the real-time processing mode can meet the requirements of data analysis and response in an edge environment.

Through the steps, the collected multi-mode data can be processed and analyzed in real time by utilizing a lightweight deep learning model and algorithm on the edge equipment, so that the high-efficiency utilization and value extraction of the data are realized. In particular, an event-driven streaming approach may be employed that processes data based on the triggering of an event. When new data arrives, the system triggers corresponding events and then executes corresponding processing logic, in multi-modal data processing, corresponding processing events can be triggered according to different types of data, and then processed and analyzed using appropriate models (in multi-modal data processing, a data stream processing pipeline can be established, wherein the data stream processing pipeline comprises a plurality of processing stages, and each stage is responsible for processing and converting different types of data, and finally outputting a result).

Step S4: and correlating the text data, the image data and the video data through a fusion model of the deep neural network, and extracting key characteristics of the drug information, wherein the key characteristics of the drug information comprise drug identification characteristics, drug text information characteristics, drug use scene characteristics and drug correlation characteristics.

Specifically, the drug identification features: identifying the appearance, package, label and other characteristics of the medicine through the image data to determine the basic information of the medicine; drug text information features: information such as names, dosage forms, components, purposes, dosages and the like of the medicaments are extracted through text data, and are key for understanding basic attributes of the medicaments. Drug use scene characteristics: and analyzing the use scene and the use mode of the medicine through the video data. Drug-related characteristics: through a fusion model of the deep neural network, text data, image data and video data are mutually related, and correlation characteristics between drug information, such as the correlation of drugs and diseases, the correlation of drugs and side effects and the like, are extracted, so that the mechanism of action and safety of the drugs are further understood.

The main purpose of extracting the key features of the drug information in step S4 described above is to reduce the dimensionality of the data and extract useful information, thereby providing a more efficient basis for the data processing and storage of the subsequent steps. This is associated with the following steps: reducing data processing and storage pressure: extracting key features of the drug information can transform the raw data into a more representative and compact form, thereby reducing the processing and storage requirements for large amounts of data in subsequent steps. This helps to reduce the computational and memory load of the edge devices, improving the efficiency and performance of the system.

Support intelligent storage management: extracting key features of the medication information may help determine which information is most important and relevant for subsequent tasks. This can provide an important basis for the deep reinforcement learning-based storage optimization algorithm in step S6, enabling it to dynamically adjust the data storage and indexing strategy according to the importance and timeliness of the drug information.

Enhancing the effectiveness of edge computing tasks: the extraction of key features of the drug information is helpful to construct a more accurate and efficient model, so that the execution efficiency and accuracy of the edge calculation task in the subsequent step are improved. This is closely related to the edge calculation collaborative optimization algorithm based on the deep reinforcement learning in step S8, and can intelligently allocate and schedule the edge calculation tasks according to the extracted features.

Therefore, the key feature of extracting the drug information in step S4 is to optimize the data processing and storage of the subsequent steps and to support efficient execution of the edge calculation task.

Step S5: based on storage resources and computing power of the edge equipment, designing a self-adaptive data storage and management strategy; further, the method comprises the steps of:

Step S5-1: first, the hardware configuration of the edge device is comprehensively identified, and the hardware configuration includes, but is not limited to, CPU type, core number, main frequency, memory capacity, hard disk type, capacity, and network bandwidth parameters, which directly affect the storage capability and computing performance of the device.

Step S5-3: counting the total amount, growth rate and proportion of different types (text, image and video) of the drug information data so as to predict future storage requirements;

Step S5-4: through log records and data analysis, the access frequency, the query mode (such as simple query, complex query, aggregate query and the like) and the read-write proportion characteristics of the drug information data are known so as to optimize the data layout and the index strategy;

Step S5-6: distributing data between a cache (e.g., RAM), a fast storage (e.g., SSD), and a mass storage (e.g., HDD) according to data access speed requirements and storage cost requirements; for example, frequently accessed hot spot data is stored in high speed media, while cold data or backup data is stored in mass media.

Specifically, the design data slicing and partitioning scheme includes:

step S5-7-1: data size and equipment capacity analysis

Data scale assessment: the total amount of current and expected pharmaceutical information data, including the amount, size, trend of growth, etc. of each category (text, image, video) of data is counted to see the overall data size.

Edge device storage capacity investigation: hardware configuration information of all edge devices is collected, and storage resources such as total capacity of memory, SSD, HDD and the like, and currently used storage space are focused. Ensuring that the data slicing and partitioning scheme does not exceed the actual storage capacity of the device.

Step S5-7-2: data slicing principle and strategy

Data fragmentation definition: data slicing (DATA SHARDING) refers to dividing a large dataset into several smaller, independent logical segments, each segment being referred to as a data slice. The main purpose of the slicing is to distribute data to a plurality of storage nodes, realize horizontal expansion and improve the performance and capacity of data storage and access.

Selecting a slicing key: a shard key (Shard Key) is determined, which is the basis for determining how data is allocated to each shard. For medication information data, possible choices are medication ID, medication category, time stamp, geographic location, etc. The selection of the slicing keys should consider factors such as the distribution uniformity, the query mode, the frequency of data addition and deletion and the like, ensure that the data distribution after slicing is as balanced as possible and are beneficial to common query operation.

Determining the number of fragments: the number of suitable slices is estimated based on the data size, the device storage capacity and the expected future growth, as well as the number of edge devices available to the system. The number of fragments can fully utilize the storage resources of the equipment, and can keep the moderate dispersion of the data distribution, so that the single fragments are prevented from being too large or too small.

Step S5-7-3: data partition design

Data partition definition: the data partition (Data Partitioning) is within a single storage node, dividing the data into logical regions according to a certain rule. The main purpose of partitioning is to optimize data access performance and management efficiency within a single storage device.

Partition strategy selection: and selecting a proper partition strategy according to the data characteristics and the query requirements. Common partitioning strategies include:

range partitioning: the partitions are divided according to one continuous attribute (such as medicine ID and time range) of the medicine information, and the method is suitable for the condition that the data query is mainly based on the interval range.

List partitioning: the method is suitable for the situation that the data query is mainly based on a fixed value list.

Hash partitioning: the hash function is used for mapping the medicine information to different partitions, so that the data is uniformly distributed, and the method is suitable for the situations that data query has no obvious preference and the data of each partition needs to be kept balanced.

Partition key setting: partition keys, i.e., attributes for performing partition operations, are determined. The partition key should be coordinated with the partition key to ensure that partition operations do not disrupt the partition's equilibrium.

Step S5-7-4: data slicing and partitioning implementation

And (3) fragment distribution: the designed data fragments are uniformly distributed to the edge devices, so that the approximate balance of the number of fragments and the data quantity carried by each device is ensured. The shard allocation may be performed using consistent hashing, polling allocation, etc. algorithms.

Partition creation: on each edge device, a corresponding partition structure is created according to the selected partition policy. For database systems, it may involve creating a new table or index; for file systems, it may involve creating a new directory structure.

Data migration and initialization: and reorganizing and migrating the existing drug information data according to the slicing and partitioning schemes to ensure that the data are correctly distributed into the corresponding slicing and partitioning. For newly added data, it should be ensured that the data is written following the sharding and partitioning rules.

Step S5-7.5: data slicing and partition maintenance

Monitoring and optimizing: the actual use condition of the data fragments and the partitions is continuously monitored, and the actual use condition comprises indexes such as data distribution, storage space utilization rate, query performance and the like. Necessary adjustment is performed according to the monitoring result, such as re-balancing the fragments, adjusting the size of the partitions, optimizing the index, and the like.

Expansion and contraction: with the increase of the data scale or the increase and decrease of the equipment, the partition and partition schemes are adjusted in time, such as the number of the partitions is increased, the size of the partitions is adjusted, the data is migrated, and the like, so as to adapt to new storage requirements and equipment resources.

Through the steps, the data slicing and partitioning scheme is designed and implemented according to the data scale and the storage capacity of the edge equipment, the medicine information data can be effectively distributed into the distributed edge computing network, storage resources are fully utilized, the data storage and access performance is improved, and high-efficiency support is provided for subsequent data processing, analysis and query operations.

Step S5-8: acquiring a data access mode analysis result, and dynamically adjusting the layout of data in a storage medium by combining the edge equipment real-time resource monitoring data;

specifically, the step of obtaining the analysis result of the data access mode includes:

Historical access record analysis: historical access logs of medication information data are collected, including access time, query type (read-write, query conditions, aggregate operations, etc.), data range involved, and the like. By statistical analysis, the rules and hot spots of data access, such as the type of drug, time period, query condition and the like of high-frequency access, are identified.

Future access prediction: and by combining with external information such as business development trend, seasonal factors, sales promotion activities and the like, reasonable prediction is carried out on future data access modes, and prospective reference is provided for dynamic adjustment.

The real-time resource monitoring of the edge equipment comprises the following steps:

Hardware resource monitoring: the CPU utilization rate, the memory occupancy rate, the disk I/O, the network bandwidth and other key performance indexes of the edge equipment are continuously monitored, and the real-time use condition of equipment resources is known.

Storage medium state monitoring: particularly, the indexes such as the utilization rate, the read-write speed, the I/O waiting time and the like of each storage medium (such as a memory, an SSD and an HDD) are concerned, and the current service capacity and the load condition of each medium are evaluated.

And calculating the heat value of each part of the drug information data according to the analysis result of the data access mode. The heat value can comprehensively consider factors such as access frequency, query complexity, data importance and the like, reflect the activity degree of the data, and divide the drug information data into hot spot data, warm data and cold data according to the data heat value. The hot spot data is data with frequent access and high response time requirement; the temperature data is data with lower access frequency but certain access requirement; cold data refers to data that is accessed little to no. Wherein dynamically adjusting the data layout comprises: hot spot data optimization: the hot spot data is stored as much as possible on a storage medium, such as a memory or SSD, where the access speed is the fastest. If the memory resources are limited, it is considered to use a cache policy such as LRU (LEAST RECENTLY Used) to keep the recently accessed hot spot data in the memory. And (3) temperature data adjustment: and determining the storage position of the temperature data on the SSD or the HDD according to the access frequency of the temperature data and the resource condition of the edge equipment. If the resources allow, part of the temperature data can be properly reserved on the SSD to cope with the possible access requirements; otherwise, it may be moved to the HDD, freeing SSD space for hotter data. Cold data migration: cold data is migrated to lower cost, higher capacity HDDs. For cold data that is not accessed for a long period of time and no longer requires immediate access, archival storage or deletion may be further considered to save storage space. After the data layout is adjusted, various performance indexes of the edge equipment, such as response time, I/O throughput, resource utilization rate and the like, are continuously monitored, the adjustment effect is verified, the data layout strategy is timely adjusted according to the monitoring result, and closed-loop management of dynamic adjustment of the data layout is formed. Through the steps, the layout of the data in the storage medium is dynamically adjusted according to the drug information data access mode and the real-time resource condition of the edge equipment, so that the data access performance is improved to the maximum extent, the storage resource utilization is optimized, and the effective management and the efficient access of the drug information data in the edge computing environment are ensured.

In summary, the method and the device can comprehensively evaluate the resource of the edge equipment and deeply analyze the characteristics of the drug information data through the step S5, design and implement the self-adaptive data storage architecture and management strategy, ensure that the drug information data is efficiently stored, quickly accessed, reliably protected and flexibly managed under the limited storage resource and calculation condition of the edge equipment, and further meet the actual service requirements.

Step S6: adopting a storage optimization algorithm based on deep reinforcement learning, and dynamically adjusting a data storage and index strategy according to the importance and timeliness of the drug information so as to realize intelligent utilization of storage resources; further, the method comprises the steps of:

Step S6-2: state variables of the storage optimization problem are defined, including importance and timeliness of current drug information data, use condition of storage resources (such as memory, SSD, HDD), data access mode and system load, and state vectors of the optimization problem are formed through the state variables.

Step S6-3: the state transition model is selected to describe the law of state change in the storage environment, such as a Markov Decision Process (MDP) or a Partially Observable Markov Decision Process (POMDP), and can reflect key characteristics such as data importance and timeliness change, dynamic allocation of storage resources, evolution of data access modes and the like.

Step S6-4: determining the action space of the storage optimization problem, namely, the possible optimization measures, such as adjusting the distribution of data on different storage media, changing an index structure, starting data compression or deduplication, and executing data cold-hot migration.

Step S6-5: constructing a deep reinforcement learning model to capture time series characteristics of data importance and timeliness; specifically, for example, deep Q Network (DQN), double DQN, policy gradient method (e.g., REINFORCE, actor-Critic), etc. may be employed. The model should include components of input layer (receiving state vector), hidden layer (performing feature learning), output layer (generating motion probability distribution or Q value), etc. Deep learning models such as Convolutional Neural Networks (CNNs), recurrent Neural Networks (RNNs), and the like may be used to capture time series characteristics of data importance and timeliness.

Step S6-6: a simulator for constructing a storage environment by utilizing historical data or synthesized data so as to generate interaction samples in a model training process comprises the following steps:

Step S6-6-1: and (3) bonus function design: according to the optimization objective, a reward function (Reward Function) is designed for measuring the revenue of the system after taking action in each state.

Step S6-6-2: model training: a deep reinforcement learning model is trained by using reinforcement learning algorithm (such as Q-learning, policy Gradient and the like). In the training process, the model learns the optimal storage strategy through interaction with the simulation environment. Experience playback, target network, exploration-utilization trade-off, etc. techniques may be employed to improve training efficiency and stability.

Step S6-7: dynamic optimization is carried out by applying a deep reinforcement learning model, and the method specifically comprises the following steps:

Step S6-7-1: real-time state sensing: state variables of the storage environment, such as data importance and timeliness, storage resource use conditions, data access modes and the like, are monitored in real time to form a current state vector.

Step S6-7-2: policy generation and enforcement: and inputting the current state vector into a trained deep reinforcement learning model to generate an optimal storage strategy (namely action). The policy is enforced, such as adjusting the data storage layout, updating the index structure, etc.

Step S6-7-3: feedback and model update: observing the actual performance of the system after executing the strategy, calculating a prize value, and adding the (state, action, prize) triples as new samples to the experience playback buffer. Samples are periodically sampled from the buffer area, and the deep reinforcement learning model is subjected to online fine adjustment, so that the strategy of the deep reinforcement learning model is continuously adapted to the change of the storage environment.

Through the steps, the storage optimization algorithm based on deep reinforcement learning is adopted, and the data storage and index strategy can be dynamically adjusted according to the importance and timeliness of the drug information, so that the intelligent utilization of storage resources is realized. The method can automatically learn and adapt to the change of the complex storage environment, continuously optimize the storage resource allocation, improve the data access performance, reduce the storage cost and ensure the data reliability.

Step S7: establishing a collaboration and optimization mechanism among distributed edge computing nodes in edge equipment to realize dynamic scheduling and collaborative computing of edge computing resources, wherein the method comprises the following steps:

step S7-1: defining edge computing node roles and communication protocols, specifically, the method comprises the following steps:

Step S7-1-1: role division: according to the hardware performance, geographic position, network condition and other factors of the edge equipment, the edge computing nodes are divided into a master node, a slave node and an agent node. The master node is responsible for global resource scheduling and task allocation, the slave node executes specific calculation tasks, and the proxy node is responsible for data forwarding and communication support.

Step S7-1-2: communication protocol setting: protocol standards for communication between edge computing nodes, such as message formats, data encoding, transmission protocols (e.g., TCP/IP, UDP), quality of service (QoS) requirements, etc., are formulated to ensure efficient and reliable information exchange.

Step S7-2: constructing a distributed collaborative framework;

Specifically, constructing the distributed collaboration framework includes:

Step S7-2-1: and constructing a task distribution system on the master node, and being responsible for receiving application requests from the cloud or local, decomposing the task into sub-tasks according to factors such as task properties, resource requirements, node states and the like, and distributing the sub-tasks to proper slave nodes for execution.

Step S7-2-2: and designing a resource scheduling algorithm adapting to edge computing characteristics, such as dynamically adjusting computing resource allocation based on policies such as load balancing, data locality, energy consumption optimization, task priority and the like. Centralized, distributed, or hybrid scheduling mechanisms may be employed.

Step S7-2-3: and establishing an edge computing node state monitoring and feedback mechanism, collecting information such as CPU utilization rate, memory occupation, network bandwidth, task execution progress and the like of the node, updating a node state database in real time, and feeding the information back to a task distribution system and a resource scheduling algorithm.

Step S7-3: the method for realizing cooperative computing specifically comprises the following steps:

Step S7-3-1: efficient synchronization and sharing of data among edge computing nodes is achieved, including the use of distributed file systems, key value storage, message queuing techniques to ensure consistency and availability of data among nodes.

Step S7-3-2: for tasks which need to be completed by a plurality of nodes in a coordinated way, mechanisms such as task decomposition, inter-task communication, result merging and the like are designed, and further, parallel computing models such as MapReduce, stream computing, graph computing and the like can be adopted.

Step S7-3-3: error-tolerant mechanisms such as node fault detection, task failure retry, data recovery and the like are realized, and the robustness and high availability of the edge computing task are ensured.

Through the steps, a collaboration and optimization mechanism between distributed edge computing nodes is established in the edge equipment, so that the dynamic scheduling and collaboration computing of edge computing resources are realized, the advantages of edge computing can be fully utilized, the data processing efficiency is improved, the network transmission load is reduced, and the service quality and the data safety are ensured.

Step S8: the intelligent allocation and the scheduling of the edge calculation tasks are realized through the edge calculation collaborative optimization algorithm based on the deep reinforcement learning, so that the overall performance and the efficiency of the system are improved; further, the edge computing collaborative optimization algorithm based on the deep reinforcement learning comprises the following steps:

step S8-1: the type and the characteristics of the edge computing task are defined, such as the computing resource requirements (CPU, memory, GPU utilization rate and the like), task priority and time constraint of the drug information processing task (such as real-time analysis, feature extraction, data fusion and the like); (execution deadline, real-time Property requirement)

Step S8-2: constructing a state model for reflecting the whole edge computing system; including but not limited to:

Edge device status: the real-time utilization rate, the residual capacity, the equipment health status and the like of hardware resources (such as a CPU, a memory, a storage space, a network bandwidth and the like) of each edge equipment are recorded.

Task queue status: the number, type, priority and respective resource requirements of the tasks to be allocated, executing and completed are recorded.

Network topology state: network conditions such as connectivity between edge devices, communication delay, etc. are described.

Step S8-3: designing a deep reinforcement learning model architecture, specifically including: step S8-3-1: state coding: the environmental states defined in step S8-1 are converted into input vectors that can be understood by the deep reinforcement learning model. The states may be encoded using structures such as a multi-layer perceptron (MLP), convolutional Neural Network (CNN), or Recurrent Neural Network (RNN).

Step S8-3-2: action space definition: define actions that the intelligent agent can take, i.e., allocation decisions for edge computing tasks. Actions may include selecting an edge device to perform a task, determining a task execution order, adjusting task priority, dynamically adjusting task partitioning (e.g., task splitting or merging), and so forth.

Step S8-3-3: and (3) bonus function design: a reward function is set to measure decision quality of the intelligent agent. The reward function should comprehensively consider factors such as task completion (e.g. task completion speed, success rate), resource utilization (e.g. avoiding resource waste, preventing overload), task priority satisfaction, system stability, etc. The reward function should encourage efficient task allocation, resource balancing utilization, and quick response to high priority tasks.

Step S8-3: training a deep reinforcement learning model, specifically, including:

Step S8-3-1: a simulation environment is built or created using existing tools (e.g., openAI Gym, etc.) to simulate the operational state and task execution of a real-edge computing system. The environment should be able to dynamically update the state based on the decisions of the intelligent agents and return corresponding reward signals.

Step S8-3-2: the intelligent agent is trained using a deep reinforcement learning algorithm (e.g., DQN, DDPG, A, C, PPO, etc.). In the training process, the intelligent agent continuously tries different task allocation strategies in the simulation environment, and updates strategy parameters according to received reward signals so as to learn the optimal or near-optimal task allocation strategies step by step.

Step S8-4: deployment and real-time decision making, specifically, include: step S8-4-1: model deployment: and deploying the trained deep reinforcement learning model to a controller or a coordination node of the actual edge computing system.

Step S8-4-2: and (3) real-time task scheduling: during actual operation, the controller collects current environmental state information periodically or on demand, and inputs the current environmental state information into the deployed deep reinforcement learning model. And outputting an optimal task allocation decision by the model, and guiding the edge computing system to perform task scheduling. According to the system state change and the addition of new tasks, the model continuously makes real-time decisions, and ensures that the tasks are distributed and executed among the edge devices efficiently and stably and meet the priority requirements.

In summary, step S8 implements intelligent allocation and scheduling of the edge calculation task by the deep reinforcement learning technology. The process comprises defining task and environment states, designing a deep reinforcement learning model architecture, model training and real-time decision after deployment, and aims to optimize edge computing resource utilization, improve task processing efficiency, guarantee task priority satisfaction and overall system stability.

Step S9: storing the drug information on a plurality of nodes in a distributed network to realize decentralization storage; such architecture allows drug information to no longer rely on a single centralized storage server, improving reliability and security of the system, including:

Step S9-1: protocol type selection: an off-centered storage protocol suitable for drug information storage, such as IPFS (InterPlanetary File System), swarm, storj, filecoin, etc., is determined. These protocols provide for decentralised content addressing, data slicing, redundant storage, version control, etc.

Step S9-2: an appropriate decentralized storage service platform is selected, such as an official client based on the selected protocol, an open source library, an SDK or API interface provided by a third party service provider. The platform should support operations such as uploading, retrieving, updating and deleting of the drug information.

Step S9-3: and (3) data packaging: and packaging the medicine information (including multi-mode data such as texts, images, videos and the like acquired in the step S2 and processed in the steps S3-S8, and related metadata such as medicine identification, time stamp, version information and the like) according to the requirements of the selected protocol.

Step S9-3: and (3) hash generation: a hash function (e.g., SHA-256) is applied to the encapsulated data block to generate a unique Content Identifier (CID). The CID serves as an address for data for locating and referencing the data block in the network.

Step S9-4: data slicing: and cutting the encapsulated drug information data block according to the size specified by the protocol to generate a plurality of smaller data fragments, wherein each fragment has an independent CID.

Step S9-5: redundancy configuration: the replication factor of the data segment is set (replication factor) according to a redundancy policy of the decentralized storage network. The replication factor determines the number of nodes each data segment should be replicated for storage in the network, and is typically used to ensure fault tolerance and availability of the data.

Step S9-6: node discovery: the available storage nodes (storage nodes are data center servers) are looked up by a mechanism provided by the protocol, such as DHT distributed hash tables.

Step S9-7: storage allocation: according to the storage allocation algorithm of the selected protocol, the data fragments and the backups thereof are allocated to different storage nodes (the allocation strategy needs to consider the factors of storage capacity, network bandwidth, geographical distribution and the like of the nodes so as to realize balanced distribution and efficient access of the data).

Step S9-8: and (3) data transmission: the data segments are sent to the assigned storage nodes over a point-to-point (P2P) network.

Step S9-9: storing confirmation: after receiving the data, the storage node checks the data (based on CID), confirms that the data is complete and correct, stores the data locally, and returns storage confirmation information to the uploading party. The uploading party gathers enough storage confirmation information to ensure that the data reaches a predetermined redundancy level.

When the medicine information changes, the new version data is subjected to the steps of packaging, hashing, slicing and the like to generate a new CID. Old version data is reserved or cleared by strategy to support historical version tracing; the online state, the storage capacity, the response speed and the like of the storage nodes are checked periodically or according to the requirement, so that the availability of the data is ensured. And if the node is found to fail or the performance of the node is reduced, triggering the data to migrate to other healthy nodes, and maintaining the redundancy level.

Step S9-10: the user or application initiates a query request to the distributed network through the CID providing the medication information, the network locates the node storing the corresponding data fragment according to the CID, and the node returns the data to the requester.

In summary, the step S9 is performed to store the drug information on a plurality of nodes in the distributed network, thereby implementing the decentralization storage. The architecture enhances the reliability and the safety of the system, reduces the single-point fault risk, and lays a foundation for the global high-efficiency access and sharing of the drug information.

Step S10: an access control strategy is defined on a distributed network through an intelligent contract, and a homomorphic encryption method is adopted to encrypt the drug information, so that only authorized users can access specific drug information. The intelligent contract can dynamically verify and authorize according to the identity, authority and access purpose of the user, and ensure that the data is only accessed by the legally authorized user, and concretely comprises the following steps:

Step S10-1: intelligent contract design and deployment: according to the sensitivity, the use scene and the compliance requirement of the drug information, a detailed access control rule is formulated, wherein the rule comprises: role rights: rights to access different types of medication information are set according to different user roles (e.g., doctor, pharmacist, medication developer). Information classification: different security levels are classified according to the sensitivity degree (such as unpublished formula and clinical data) of the drug information, and different access rights are corresponding. Access conditions: and setting preconditions for accessing the drug information, such as signing a secret agreement and completing identity authentication. Operation authority: and defining the read, write, modify, delete and share operation authority of the user on the drug information. Encoding the access control policy into an executable smart contract using a smart contract programming language (e.g., solidity), the smart contract comprising the following functions: and (3) authority verification: and checking the user identity, the role and the authority state information, and judging whether specific operation is allowed. Rights management: a functional interface is provided to add or delete the checking authority, allowing the administrator to dynamically adjust the user authority. Recording an event: and recording all access requests and authority change operation logs for audit trails. The written intelligent contracts are deployed to the selected distributed network, so that the intelligent contracts become a part of the network which cannot be tampered, and the transparency and the execution force of the access control rules are ensured.

Step S10-2: according to the type, calculation requirement, performance requirement and safety standard of the drug information, the homomorphic encryption algorithm is selected, and specifically any encryption algorithm selected From Homomorphic Encryption (FHE), partial homomorphic encryption (such as RSA and ElGamal) or homomorphic encryption schemes based on grids (such as BGV and CKKS) can be adopted; the encryption library supporting the selected homomorphic encryption algorithm is selected or developed and integrated into the pharmaceutical information processing system. Ensuring that the encryption library is compatible with intelligent contracts, distributed storage systems, and the like.

Step S10-3: and (3) encrypting the medicine information by using a homomorphic encryption algorithm, generating a public key, a private key and key parameters required by sealing, generating and properly keeping encryption, wherein the public key is used for encryption, and the private key is used for decryption.

Step S10-3: storing the encrypted drug information on a plurality of nodes in a distributed network according to the distributed storage strategy of the step S9, so as to ensure that the intelligent contract can effectively manage the access of the encrypted information through the access control rule. Before accessing the drug information, the user needs to verify the identity through an identity authentication system, acquire an authorization credential (such as a digital signature, an access token and the like), and initiate an access request to the intelligent contract with the authorization credential. The smart contract verifies the access rights of the user according to a predefined access control policy, and for users that pass the rights verification, the smart contract allows them to download encrypted medication information. The user decrypts the ciphertext using the corresponding private key (or temporarily provided by the smart contract) to obtain the original drug information.

In summary, the method and the device define the access control policy, select and implement the homomorphic encryption scheme, encrypt the drug information, store the encrypted information, authorize access and use in decryption through the intelligent contract in step S10, ensure that only authorized users can access specific drug information, and effectively protect the privacy and security of the drug information.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the same; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those skilled in the art that: the technical scheme described in the foregoing embodiments can be modified or some or all of the technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit of the invention.

Claims

1. The medicine information storage method based on the distributed edge calculation and the multi-mode data is characterized by comprising the following steps:

step S6: adopting a storage optimization algorithm based on deep reinforcement learning, and dynamically adjusting data storage according to the importance and timeliness of the drug information;

in the step S6, a storage optimization algorithm based on deep reinforcement learning is adopted, and according to the importance and timeliness of the drug information, dynamically adjusting the data storage includes:

Step S6-4: determining an action space for storing the optimization problem;

step S6-5: constructing a deep reinforcement learning model to capture time series characteristics of data importance and timeliness;

Step S6-6: constructing a simulator of the storage environment by utilizing historical data or synthesized data so as to generate an interaction sample in the model training process;

2. The method for storing medication information based on distributed edge computing and multimodal data according to claim 1, wherein selecting a lightweight, low power consumption embedded system suitable for edge computing as an edge device in step S1 comprises:

Step S1-1: demand analysis:

Step S1-2: edge device selection:

3. The method for storing medication information based on distributed edge computing and multimodal data according to claim 2, wherein the customizing the operating system and software stack of the device in step S1 comprises:

Step S1-3: and (3) customizing design:

step S1-4: performance evaluation and testing:

step S1-5: optimizing and adjusting:

4. The method for storing drug information based on distributed edge computing and multi-modal data according to claim 1, wherein the establishing the multi-modal data fusion and processing module on the edge device in step S3 includes:

Step S3-1: determining a resource limit of the edge device:

Step S3-2: analyzing the multimodal data types:

step S3-3: designing a data input interface:

Step S3-4: data preprocessing:

Preprocessing input multi-mode data to reduce noise and improve data quality;

Step S3-5: multimodal data fusion:

5. The method for storing medication information based on distributed edge computing and multi-modal data according to claim 4, wherein the selecting a lightweight deep learning model for real-time processing and analysis of the acquired multi-modal data comprises:

Step S3-7: selecting a lightweight deep learning model:

step S3-8: model deployment and optimization:

step S3-9: real-time processing and analysis:

6. The method for storing medication information based on distributed edge computing and multimodal data according to claim 1, wherein the key features of the medication information in step S4 include a medication identification feature, a medication text information feature, a medication usage scenario feature, and a medication correlation feature.

7. The method for storing medication information based on distributed edge computing and multi-modal data according to claim 1, wherein the designing of the adaptive data storage and management policy based on the storage resources and computing power of the edge device in step S5 includes:

8. The method for storing drug information based on distributed edge computing and multimodal data according to claim 1, wherein establishing a collaboration and optimization mechanism between distributed edge computing nodes in the edge device in step S7 comprises:

step S7-1: defining edge computing node roles and communication protocols;

step S7-2: constructing a distributed collaborative framework;

step S7-3: and realizing cooperative calculation.

9. The method for storing pharmaceutical information based on distributed edge computing and multi-modal data according to claim 1, wherein the edge computing collaborative optimization algorithm based on deep reinforcement learning in step S8 comprises:

step S8-3: designing a deep reinforcement learning model framework;

step S8-3: training a deep reinforcement learning model;

step S8-4: deployment and real-time decision making.