Knowledge graph-based sesri device management architecture and method

Abstract

The application discloses a SESRI equipment management architecture and method based on a knowledge graph, and the equipment management architecture comprises a data storage layer, a knowledge extraction layer, a knowledge representation layer and an application display layer. The data storage layer is responsible for providing original equipment data support for the knowledge extraction layer; the knowledge extraction layer is responsible for extracting entity and relationship structured information, and providing equipment key information for the knowledge representation layer; the knowledge representation layer is responsible for integrating the entity and relationship extracted by the knowledge extraction layer into a form of a triple, selecting a Neo4j graph database for storage, and displaying knowledge in the form of a graph; and the application display layer is responsible for providing a man-machine interactive interface, and realizing data visualization. The application realizes SESRI equipment management, solves the problems of a large number of experimental equipment in the SESRI and miscellaneous equipment data, slow equipment information retrieval speed, and difficult expression of the correlation between devices in the experimental process.

Description

Technical Field

[0001] This invention belongs to the field of equipment management and relates to an equipment management architecture and method for a Space Environment Simulation and Research Infrastructure (SESRI), specifically a SESRI equipment management architecture and method based on knowledge graphs. Background Technology

[0002] As human exploration of space deepens, there is an urgent need for in-depth research into the space environment effects on spacecraft materials, devices, and their functional systems. The space environment has a significant impact on human space activities. SESRI is a major scientific and technological infrastructure (referred to as a large-scale scientific project) designed to adapt to future large-scale space activities and technological development in my country, and to conduct basic research and engineering experiments on the interaction between the space environment and matter. SESRI focuses on major fundamental scientific and technological issues in the aerospace field, constructing my country's first large-scale research base in the field of integrated space environment and the interaction between spacecraft, living organisms, and plasma, forming an internationally leading experimental research platform for space environment coupling effects.

[0003] SESRI devices are characterized by their large scale, high cost, complex maintenance, and intricate interconnections. In the event of a malfunction, quickly obtaining basic device information, understanding inter-device connections, and pinpointing the location of the fault are crucial. SESRI device data files are numerous and mostly document-based, lacking intuitiveness and offering slow retrieval speeds. Traditional device management solutions based on management control lists cannot meet SESRI's functional requirements. Furthermore, SESRI includes a large number of experimental devices with complex interconnections. The breadth and depth of the SESRI network diagram cannot be directly observed through topology maps, making it difficult to quickly locate faulty devices manually. Therefore, researching an efficient and suitable SESRI device management solution and designing a SESRI device management platform based on this solution is of great significance for the full lifecycle management and service support of SESRI devices. Summary of the Invention

[0004] To address the challenges of managing SESRI devices and resolving issues such as slow device information retrieval and difficulty in representing relationships between devices during experiments due to the large number of experimental devices and complex device data, this invention provides a knowledge graph-based SESRI device management architecture and method. This invention designs a device management architecture with a data storage layer, a knowledge extraction layer, a knowledge representation layer, and an application display layer. Based on this architecture, an entity extraction algorithm based on active learning and BERT-BiLSTM-CRF is designed. This method utilizes an active learning sample selection strategy to select high-quality unlabeled samples and uses BERT-BiLSTM-CRF to extract device entities, improving the device entity recognition performance. Simultaneously, a relation extraction algorithm based on a dual attention mechanism is designed to reduce noise interference in extracting complex relationships between devices, improving the accuracy of the relation extraction algorithm. Finally, the triples obtained from entity and relation extraction are imported into the Neo4j graph database, and a human-computer interaction interface is designed using HTML, CSS, and JavaScript. Echarts is used to display the data visualization charts, achieving data visualization.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A SESRI device management architecture based on knowledge graphs includes a data storage layer, a knowledge extraction layer, a knowledge representation layer, and an application presentation layer. The data storage layer comprises:

[0007] The data storage layer is responsible for providing the knowledge extraction layer with raw data support from the device.

[0008] The knowledge extraction layer is responsible for extracting structured information such as entities and relationships from the raw device data provided by the data storage layer, and providing key device information for the knowledge representation layer.

[0009] The knowledge representation layer is responsible for integrating the entities and relations extracted by the knowledge extraction layer into triples, storing them in the Neo4j graph database, and displaying the knowledge in the form of a graph.

[0010] The application presentation layer is responsible for providing the human-computer interaction interface. It uses Py2neo to call knowledge representation and Echarts to display intuitive, vivid, interactive, and customizable data visualization charts to achieve data visualization.

[0011] A method for knowledge graph-based SESRI device management using the above-described device management architecture includes the following steps:

[0012] Step S101: Acquire knowledge in the SESRI equipment domain, such as equipment names, connection methods between devices, key equipment indicators, etc.

[0013] Step S102: Define the entities and their relationships in the SESRI device knowledge graph ontology, and establish the SESRI device ontology model.

[0014] Step S103: Use OWL (OntologyWed Language) to build the SESRI device ontology model;

[0015] Step S104: Select high-quality unlabeled samples using an uncertainty-based sample selection strategy;

[0016] Step S105: Use the traditional Chinese dataset entity annotation method BIOES to annotate the selected high-quality unannotated samples and construct an entity recognition corpus;

[0017] Step S106: Extract device entities using a transfer learning algorithm based on BERT-BiLSTM-CRF;

[0018] Step S107: Use the SPO triplet form to extract relations from the device information and build the original corpus for the relation extraction experiment;

[0019] Step S108: Use a dual attention mechanism algorithm to extract the relationships between device entities;

[0020] Step S109: Import the extracted triples into the Neo4j graph database to represent the knowledge in the form of a graph;

[0021] Step S110: Use HTML, CSS and JavaScript to design the human-computer interaction interface, and use Echarts to display intuitive, vivid, interactive and customizable data visualization charts to achieve data visualization.

[0022] Compared with the prior art, the present invention has the following advantages:

[0023] 1. Knowledge graphs can display complex domain knowledge through data mining, information processing, and other methods, and explain the development law of domain knowledge. This characteristic can solve the problem of the difficulty in expressing the relationship between various devices in the experimental process of complex devices in SESRI.

[0024] 2. Utilizing knowledge graphs to construct a knowledge base for complex devices can comprehensively encompass the basic information of each device in a complex device, the connection methods between devices, and the key indicators of the devices. Presented in a graph structure, it can better express the complexity and hierarchy of device connections in SESRI, assisting in quickly locating the location of faults and providing a more efficient and suitable SESRI device management method for large-scale scientific projects. Attached Figure Description

[0025] Figure 1 A block diagram of the SESRI device management architecture based on knowledge graphs;

[0026] Figure 2 For the SESRI device knowledge graph ontology model;

[0027] Figure 3 This is a flowchart of SESRI device management based on knowledge graphs. Detailed Implementation

[0028] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.

[0029] This invention provides a knowledge graph-based SESRI device management architecture, such as... Figure 1 As shown, the device management architecture consists of four parts: a data storage layer, a knowledge extraction layer, a knowledge representation layer, and an application presentation layer. The data storage layer comprises:

[0030] The data storage layer is responsible for providing the knowledge extraction layer with raw data support from the device.

[0031] The knowledge extraction layer is responsible for extracting structured information such as entities and relationships from the semi-structured and unstructured data in the data storage layer, providing key device information for the knowledge representation layer.

[0032] The knowledge extraction layer is responsible for extracting structured information such as entities and relationships from the semi-structured and unstructured data in the data storage layer, providing key device information for the knowledge representation layer.

[0033] The knowledge representation layer is responsible for integrating the entities and relations extracted by the knowledge extraction layer into triples, storing them in the Neo4j graph database, and displaying the knowledge in the form of a graph to enhance the computer's ability to retrieve information and improve its intuitiveness.

[0034] The application presentation layer is responsible for providing the human-computer interaction interface. It uses Py2neo to call knowledge representation, and uses HTML, CSS and JavaScript to write and design the human-computer interaction interface. It uses Echarts to display intuitive, vivid, interactive and customizable data visualization charts to realize data visualization.

[0035] In this invention, the original data files of the devices are stored in formats such as Word, PDF, and TXT, and the file content includes information such as device list, device bitmap, device interface details, and device ID information table.

[0036] In this invention, the data storage layer uses Ceph as the storage medium to perform distributed storage of the device's original data files.

[0037] In this invention, the knowledge extraction layer consists of three parts: device knowledge graph ontology, entity extraction, and relation extraction, wherein:

[0038] The device knowledge graph ontology is a specification for modeling device concepts, providing a clear definition of device concepts and the relationships between them in a formal way;

[0039] The device knowledge graph ontology consists of two parts: entities and their relationships.

[0040] The entities comprise eight categories: "Sys" (system or subsystem), "Subsys" (subsystem), "Equip" (device), "L" (device location), "ID" (device ID), "Brand" (brand and model), "Indicator" (metric), and "Interface" (interface). Specifically: "Sys" describes the name of the system or subsystem to which the device belongs; "Subsys" describes the name of the subsystem under the system or subsystem to which the device belongs; "Device Name" describes the name of the device; "Device Location" describes the specific location of the device in SESRI; "Device ID" describes the ID of the device; "Brand and Model" describes the brand and specific model of the device; "Key Indicator" describes the key indicator information of the device; and "Interface" describes the interface information of the device.

[0041] The relationships described here are used to describe the relationship types between entities, and there are a total of 8 relationship categories, namely ["Sys" (system or subsystem), "Include", "Subsys" (subsystem)], ["Subsys" (subsystem), "Include", "Equip" (device)], ["Equip" (device), "ID_of" (identifier), "ID" (device ID)], ["Equip" (device), "Indicator_of" (indicator), "Indicator" (indicator)], ["Equip" (device), "Include", "Interface" (interface)], ["Equip" (device), "Type_of" (model), "Brand" (brand model)], ["Equip" (device), "Locate", "L" (device location)], ["Interface", "Connection", "Interface" (interface)];

[0042] The entity extraction refers to automatically identifying named entities, such as device and device ID, from the text dataset in the data storage layer.

[0043] The entity extraction uses the BERT-BiLSTM-CRF transfer learning algorithm to address the issue of a small number of device-labeled samples in the text dataset of the data storage layer.

[0044] The relation extraction refers to extracting the relationships between entities from the text dataset of the data storage layer, and linking the entities obtained by entity extraction through the relationships.

[0045] The relationship extraction employs a dual attention mechanism algorithm to address the problem of complex inter-device relationships in text datasets at the data storage layer, which makes entity relationship identification difficult.

[0046] A method for knowledge graph-based SESRI device management utilizing the aforementioned device management architecture, such as... Figure 3 As shown, the method includes the following steps:

[0047] Step S101: Acquire knowledge in the SESRI equipment domain, such as equipment names, connection methods between devices, key equipment indicators, etc.

[0048] Step S102: Define the entities and their relationships in the SESRI device knowledge graph ontology, and establish the SESRI device ontology model.

[0049] Step S103: Use OWL (OntologyWed Language) to build the SESRI device ontology model;

[0050] Step S104: Select high-quality unlabeled samples using an uncertainty-based sample selection strategy;

[0051] Step S105: Use the traditional Chinese dataset entity annotation method BIOES to annotate the selected high-quality unannotated samples and construct an entity recognition corpus;

[0052] Step S106: Extract device entities using a transfer learning algorithm based on BERT-BiLSTM-CRF;

[0053] Step S107: Use the SPO triplet form to extract relations from the device information and build the original corpus for the relation extraction experiment;

[0054] Step S108: Use a dual attention mechanism algorithm to extract the relationships between device entities;

[0055] Step S109: Import the extracted triples into the Neo4j graph database to represent the knowledge in the form of a graph;

[0056] Step S110: Use HTML, CSS and JavaScript to design the human-computer interaction interface, and use Echarts to display intuitive, vivid, interactive and customizable data visualization charts to achieve data visualization.

[0057] Example 1:

[0058] This embodiment provides a SESRI device management architecture based on a knowledge graph, such as... Figure 1 As shown, the device management architecture includes a data storage layer, a knowledge extraction layer, a knowledge representation layer, and an application presentation layer. The data storage layer comprises:

[0059] The data storage layer is responsible for providing the knowledge extraction layer with raw data support from the device.

[0060] The knowledge representation layer is responsible for integrating the entities and relations extracted by the knowledge extraction layer into triples, storing them in the Neo4j graph database, and displaying the knowledge in the form of a graph to enhance the computer's ability to retrieve information and improve its intuitiveness.

[0061] The application presentation layer primarily provides a human-computer interaction interface. It uses Py2neo to call knowledge representations, adopts the Flask architecture, and uses HTML, CSS, and JavaScript to design and write the human-computer interaction interface. It provides user access and device query functions, and uses Echarts to display intuitive, vivid, interactive, and customizable data visualization charts to achieve data visualization.

[0062] In this embodiment, the original device data files are stored in formats such as Word, PDF, and TXT. The file content includes information such as a device list, device bitmap, detailed interface table of the digital simulation control system, device ID information table, and device user manual. The data storage layer uses Ceph as the storage medium for unstructured data to perform distributed storage of device data information.

[0063] In this embodiment, the knowledge extraction layer consists of three parts: device knowledge graph ontology, entity extraction, and relation extraction. The device knowledge graph ontology describes identification knowledge (such as device names included in SESRI), relational knowledge (such as the connection methods and hierarchical relationships of devices in SESRI), and key indicators of devices.

[0064] In this embodiment, as Figure 2 As shown, the device knowledge graph ontology comprises two parts: entities and their relationships, wherein:

[0065] The entities comprise eight main categories: "Sys" (system or subsystem), "Subsys" (subsystem), "Equip" (equipment), "L" (equipment location), "ID" (equipment ID), "Brand" (brand / model), "Indicator" (indicator), and "Interface" (interface). The "System or Subsystem" describes the name of the system or subsystem to which the equipment belongs, specifically including numerical simulation control system, integrated subsystem, life science subsystem, device subsystem, and microscopic subsystem. The "Subsystem" describes the name of the subsystems within the system or subsystem to which the equipment belongs, specifically including numerical simulation analysis subsystem, system-level integrated irradiation test chamber, microbeam irradiation cell microscopy subsystem, ion irradiation terminal integrated system, and space... The equipment names describe the names of the equipment, specifically including Experimental System Access Switch-1, Local Data Acquisition Switch-1, etc.; the equipment locations describe the specific locations of the equipment within SESRI, specifically including Cabinet-112U in the Multifunctional Room of the Life Subsystem on the 2nd Floor of the Comprehensive Building, Cabinet-26U in the Operation Room of the Life Subsystem on the 3rd Floor of the Comprehensive Building, etc.; the equipment IDs describe the IDs of the equipment, specifically including ZH-XTJZHFZSYC-ZHCRQXT-null-null-1, SM-WSFZXBXWZXT-WSFZSGZGZZ-WSFZSGZXWJ-null-1, etc.; the brand and model descriptions describe the brand and specific model of the equipment, with specific examples including H3C. S-5170-36F-EI, H3CLS-5130S-28TP-EI, etc.; the key indicators describe the key indicator information of the equipment, and specific examples include switching capacity, packet forwarding rate, total number of interfaces, etc.; the interfaces describe the interface information of the equipment, and specific examples include 1Gbps electrical ports, 1Gbps optical ports, etc.

[0066] The relationship describes the relationship types between entities, which include a total of 8 relationship categories, namely ["Sys" (system or subsystem), "Include", "Subsys" (subsystem)], ["Subsys" (subsystem), "Include", "Equip" (device)], ["Equip" (device), "ID_of", "ID" (device ID)], ["Equip" (device), "Indicator_of", "Indicator" (indicator)], ["Equip" (device), "Include", "Interface" (interface)], ["Equip" (device), "Type_of", "Brand" (brand model)], ["Equip" (device), "Locate", "L" (device location)], ["Interface", "Connection", "Interface" (interface)].

[0067] In this embodiment, entity extraction refers to automatically identifying named entities, such as devices and device IDs, from the text dataset in the data storage layer. Entity extraction employs the BERT-BiLSTM-CRF transfer learning algorithm to address the issue of a small number of labeled device samples in the text dataset of the data storage layer.

[0068] In this embodiment, relation extraction refers to extracting the relationships between entities from the text dataset in the data storage layer, and linking the extracted entities together through these relationships. Relationship extraction employs a dual attention mechanism algorithm to address the problem of complex inter-device relationships in the text dataset of the data storage layer, which makes entity relationship identification difficult.

[0069] Example 2:

[0070] This embodiment provides a SESRI device management method based on knowledge graphs, such as... Figure 3 As shown, the flow of the method is as follows:

[0071] Step S101: Obtain knowledge in the SESRI equipment field, such as equipment names, connection methods between equipment, key equipment indicators, etc.

[0072] In this embodiment, knowledge of the device domain, such as device names like cabinets, monitoring servers, monitors, and data and management switches, is obtained from SESRI device engineering documents.

[0073] Step S102: Define the entities and their relationships in the SESRI device knowledge graph ontology, and establish the SESRI device ontology model.

[0074] In this embodiment, the entities comprise eight categories: "Sys" (system or subsystem), "Subsys" (subsystem), "Equip" (device), "L" (device location), "ID" (device ID), "Brand" (brand / model), "Indicator" (indicator), and "Interface" (interface); the relationships comprise eight types: ["Sys" (system or subsystem), "Include", "Subsys" (subsystem)], ["Subsys" (subsystem), "Include", "Equip" (device)], ["Equip" (device), "I" (device location), "I" (device ID ... [D_of (identifier), "ID (device ID)], ["Equip (device), "Indicator_of", "Indicator" (indicator)], ["Equip (device), "Include", "Interface" (interface)], ["Equip (device), "Type_of", "Brand" (brand model)], ["Equip (device), "Locate", "L" (device location)], ["Interface", "Connection", "Interface" (interface)].

[0075] Step S103: Use OWL (Ontology Web Language) to build the SESRI device ontology model.

[0076] Step S104: Select high-quality unlabeled samples using an uncertainty-based sample selection strategy.

[0077] In this embodiment, CRF is used as the model to provide the posterior probability of sequence samples. The uncertainty of the samples is measured by calculating the sequence confidence of unlabeled samples. Sample selection is performed based on the principle of minimum confidence, selecting sequence samples with high sequence confidence values ​​in batches, and then selecting samples with high information density from them. The specific selection method is as follows:

[0078] (1) Calculate φ LC (x)=1-P(y) * |x)

[0079] (2) Calculation

[0080] (3) Calculation

[0081] (4) Calculate φ ID (x)=φ SE (x)·φSR (x)

[0082] Among them, y * It is the most likely label sequence corresponding to sequence sample x, P(y) * |x) represents the confidence level of the corresponding sequence sample, φ LC (x) represents the sequence confidence score, T represents the length of the sequence sample, M represents the total number of labels, m represents the number of labels, and P(y) represents the confidence score of the sequence. t =m) represents the marginal probability of the character at position t in the sequence based on the prediction results of the current model, U is the number of other samples in the unlabeled sample pool, and x (u) sim(x, x) represents the u-th sample in the unlabeled sample pool. (u) ) represents sample x and sample x (u) The similarity, φ SR (x) represents the representativeness of sample x, φ SE (x) represents the information content of sample x, φ ID (x) Information density of sample x.

[0083] Step S105: Use the traditional Chinese dataset entity annotation method BIOES to annotate the selected high-quality unannotated samples and construct an entity recognition corpus.

[0084] In this embodiment, the traditional Chinese dataset entity annotation method BIOES (B-begin, I-inside, O-outside, E-end, S-single) is used to construct the device management dataset. Each character in the device management dataset corresponds to a label, which is used as the extraction result. Here, B (Begin) indicates that the character is at the beginning of an entity; I (Inside) indicates inside; O (Outside) indicates outside; E (End) indicates that the character is at the end of an entity; and S (Single) indicates a single character, which is itself an entity. BIOES annotation labels each element as "BX", "IX", "EX", "SX", or "O", where "X" represents a noun phrase (NP), often indicating the entity category of the dataset.

[0085] Step S106: Extract device entities using a transfer learning algorithm based on BERT-BiLSTM-CRF.

[0086] In this embodiment, the BERT-BiLSTM-CRF transfer learning algorithm consists of three layers: BERT, BiLSTM, and CRF. First, the pre-trained BERT language model is used to initialize and obtain character vectors from the input text, denoted as a sequence. These character vectors effectively extract features from the text by utilizing the relationships between words. Second, the n-dimensional character vectors obtained by the BERT language model are used as inputs to each time step of the bidirectional long short-term memory neural network, resulting in the hidden state sequence of the bidirectional LSTM layer. (indicating forward) and (Representing backward processing), after both forward and backward processing are completed, the hidden state sequences are concatenated according to their positions to obtain the complete hidden state sequence. Then, the linear output layer maps the complete hidden state sequence to s dimensions (s is the number of label categories in the annotation set). Let the extracted sentence features be the sequence after all mappings, and let matrix L = (l1, l2, ..., l...). n )∈R n×s , l i ∈R s Each dimension l i,j Each corresponds to its character x i For each category label y i The CRF calculates the score; finally, it uses the predicted output sequence of the BiLSTM model to find the sequence that optimizes the objective function, obtains the global optimal sequence, and gets the category label value corresponding to each character.

[0087] Step S107: Use the SPO triplet form to extract relations from the device information and build the original corpus for the relation extraction experiment.

[0088] In this embodiment, based on common entity relationships in SESRI, the device information and connections between devices in various systems or subsystems of SESRI are defined using common predicate relationships, forming SPO triples of [Entity 1, Relationship, Entity 2], or [Subject, Predicate, Object]. Due to the writing style of device documentation in SESRI, the one-to-one relationship between sentences and SPO triples in the current general domain dataset does not hold; a sentence usually contains multiple relationship information. For example, in the sentence "Distributed storage cluster node-6 device interface 1 is connected to storage interconnect switch-1 interface 6," the SPO could be [Distributed storage cluster node-6, Interface, Device interface 1] or [Device interface 1, Connection, Interface 6], etc., a one-to-many SPO triple relationship.

[0089] Step S108: Use a dual attention mechanism algorithm to extract the relationships between device entities.

[0090] In this embodiment, the dual attention mechanism algorithm comprises five layers: BERT, BiLSTM, word attention mechanism, sentence attention mechanism, and CRF. First, the text is input into BERT to obtain a vector representation that integrates word vectors, segment vectors, and position vectors to capture the full-text semantic information. Second, BiLSTM extracts contextual features, taking the vector representation of the full-text semantic information as input and outputting a feature vector (l1, l2, ..., l...). n This process yields more comprehensive semantic information. Subsequently, the dual attention mechanism multiplies the feature vector output by BiLSTM with the word weight parameter and the sentence weight parameter, respectively, to obtain the output O of the dual attention mechanism. Finally, CRF adds constraints based on the output O of the dual attention mechanism to output the globally optimal label sequence.

[0091] Step S109: Import the extracted triples into the Neo4j graph database to represent the knowledge in the form of a graph.

[0092] In this embodiment, the entities and relations obtained from entity extraction and relation extraction are processed and converted into RDF format, and then imported into the Neo4j graph database to present the knowledge in the form of a graph.

[0093] Step S110: Use HTML, CSS and JavaScript to design the human-computer interaction interface, and use Echarts to display intuitive, vivid, interactive and customizable data visualization charts to achieve data visualization.

[0094] In this embodiment, Neo4j is called through the py2neo library of Python, and the device management knowledge graph is constructed using Echarts combined with the Django architecture. The human-computer interaction interface is designed using HTML, CSS and JavaScript to realize the functions of user login and device information query.

Claims

1. A method for knowledge graph-based SESRI device management, characterized in that The method comprises the following steps: Step S101: acquire knowledge in the field of SESRI equipment; Step S102: define entities and their relationships of the SESRI equipment knowledge graph ontology, and establish an SESRI equipment ontology model; Step S103: use OWL to build the SESRI equipment ontology model; Step S104: select high-quality unlabeled samples based on an uncertainty-based sample selection strategy, wherein: use CRF as a model for providing sequence sample posterior probability, measure the uncertainty of the sample by calculating the sequence confidence of the unlabeled sample, select samples based on the lowest confidence principle, select sequence samples with high sequence confidence value in batches, and select samples with high information density from them, and the specific selection method is as follows: (1) Calculate (2) Calculate (3) Calculate (4) Calculate wherein, is a sequence sample the corresponding most likely label sequence, is the corresponding sequence sample confidence, is the sequence confidence, denotes the length of the sequence sample, is the number of all labels, is the number of labels, denotes the edge probability of the character at position in the sequence according to the prediction result of the current model, is the number of other samples in the unlabeled sample pool, denotes the sample in the unlabeled sample pool, is the similarity of sample to sample , is the representation of sample , represents the information content of sample , the information density of sample ; Step S105: use the traditional entity annotation method BIOES of Chinese data set to annotate the selected high-quality unlabeled samples, and construct an entity recognition corpus; Step S106: adopt a transfer learning algorithm based on BERT-BiLSTM-CRF to extract equipment entities; Step S107: use SPO triple form to extract equipment information, and build an original corpus for relation extraction experiment; Step S108: adopt a double attention mechanism algorithm to extract the relationship between equipment entities; Step S109: import the extracted triple into the Neo4j graph database, and display the knowledge in the form of a graph; Step S110: use HTML, CSS and JavaScript to design a human-computer interaction interface, use Echarts to display intuitive, vivid, interactive and customized data visualization charts, and realize data visualization.

2. The method of knowledge graph based SESRI device management according to claim 1, wherein In step S105, the BIOES is used to construct a data set for equipment management, and each word in equipment management corresponds to a label as the extraction result, wherein B represents the beginning of an entity, I represents the inside, O represents the outside, E represents the end of an entity, and S represents a single character; the BIOES annotation marks each element as "B-X", "I-X", "E-X", "S-X" and "O", wherein "X" represents a noun phrase.

3. The method of knowledge graph based SESRI device management according to claim 1, wherein In step S106, the specific steps of the BERT-BiLSTM-CRF transfer learning algorithm are as follows: First, the pre-trained BERT language model is used to initialize and obtain character vectors from the input text information, denoted as a sequence. The obtained character vectors can effectively extract features from the text by utilizing the relationships between words; second, the BERT language model obtains... The dimensional vector is used as the input to each time step of the bidirectional long short-term memory neural network to obtain the hidden state sequence of the bidirectional LSTM layer. and , Indicates forward, This indicates a backward pass. After both the forward and backward passes are completed, the hidden state sequences are concatenated according to their positions to obtain a complete hidden state sequence. Then, the linear output layer maps the complete hidden state sequence to... dimension, Let d be the number of label categories in the annotation set, and let denote the extracted sentence features as the sequence after all mappings, and let be the matrix. Each dimension Corresponding to its characters Corresponding to each category label The CRF calculates the score; finally, it uses the predicted output sequence of the BiLSTM model to find the sequence that optimizes the objective function, obtains the global optimal sequence, and gets the category label value corresponding to each character.

4. The method of knowledge graph based SESRI device management according to claim 1, wherein In the step S108, the specific steps of the double attention mechanism algorithm are as follows: first, inputting the text into the BERT to obtain the vector representation of the full text semantic information fused with the word vector, the section vector and the position vector; second, the BiLSTM obtains the context features, the input is the vector representation of the full text semantic information fused with the word vector, the section vector and the position vector, and the output is the feature vector , to obtain more comprehensive semantic information; then, the double attention mechanism multiplies the feature vector output by the BiLSTM with the word weight parameter and the sentence weight parameter respectively to obtain the output O of the double attention mechanism. Finally, CRF adds constraints according to the output O of the double attention mechanism, and outputs the globally optimal label sequence.

5. A knowledge graph based SESRI device management architecture implementing the method of any one of claims 1-4, characterized in that The equipment management architecture comprises a data storage layer, a knowledge extraction layer, a knowledge representation layer and an application display layer. The data storage layer is responsible for providing equipment raw data support for the knowledge extraction layer; The knowledge extraction layer is responsible for extracting entity and relationship structured information from the equipment raw data provided by the data storage layer, and providing equipment key information for the knowledge representation layer; The knowledge representation layer is responsible for integrating the entities and relationships extracted by the knowledge extraction layer into the form of triples, selecting a Neo4j graph database for storage, and displaying the knowledge in the form of a graph; The application display layer is responsible for providing a human-computer interaction interface, calling the knowledge representation by using Py2neo, using Echarts to display intuitive, vivid, interactive and customized data visualization charts, and realizing data visualization.

6. The knowledge graph based SESRI device management architecture of claim 5, wherein The data storage layer uses Ceph as a storage medium to store the device original data files in a distributed manner.

7. The knowledge graph based SESRI device management architecture of claim 5, wherein The knowledge extraction layer is composed of three parts, namely, the device knowledge graph ontology, entity extraction, and relation extraction. The device knowledge graph ontology is a specification for modeling device concepts, which gives a clear definition of device concepts and their relationships in a formalized manner. The entity extraction refers to automatically identifying named entities from the text data set of the data storage layer. The relation extraction refers to extracting the association relationships between entities from the text data set of the data storage layer and connecting the entities extracted by entity extraction through relationships.

8. The knowledge graph based SESRI device management architecture of claim 7, wherein The device knowledge graph ontology includes entities and their relationships, wherein: The entities include eight categories, namely, system or subsystem, subsystem, device, device location, device ID, brand model, index, and interface. The relationships include eight relationship categories, namely, [system or subsystem, contains, subsystem], [subsystem, contains, device], [device, identification, device ID], [device, index, index], [device, contains, interface], [device, model, brand model], [device, location, device location], and [interface, connection, interface].

9. The knowledge graph based SESRI device management architecture of claim 7, wherein The entity extraction uses the BERT-BiLSTM-CRF transfer learning algorithm to solve the problem of insufficient device annotation samples in the text data set of the data storage layer. The relation extraction uses a double attention mechanism algorithm to solve the problem of complex relationships between devices in the text data set of the data storage layer, which leads to difficulties in entity relationship recognition.

Images