Centralized control scheduling service knowledge graph construction method based on machine learning

Abstract

The invention discloses a centralized control scheduling business knowledge graph construction method based on machine learning, which comprises the following steps of: extracting entity triple data related to equipment, and converting the entity triple data into a graph database to realize construction of an equipment entity graph; carrying out automatic labeling on the relationship between entities by taking classification abnormity emergency disposal and anti-error rule unstructured data as input and being assisted by the constructed equipment entity map, and carrying out weak supervision machine learning on the basis to complete the construction of a concept map; mining associated information of the operation ontology and the state ontology through machine learning, and further fusing the equipment entity map to complete construction; a training sample is expressed through feature vectors by adopting an event extraction method based on machine learning, a classification model is trained through a supervised learning method, then event instances are extracted from unlabeled texts, and construction of a case graph is completed. The construction difficulty of the knowledge graph is effectively reduced, and the accuracy is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of centralized control and dispatching, and particularly relates to a method for constructing a knowledge graph of centralized control and dispatching services based on machine learning. BACKGROUND

[0002] At present, in actual work, in terms of abnormal emergency disposal, computer monitoring systems mainly monitor and analyze, and the decision-making and execution links mostly rely on the experience accumulation and adaptability of remote centralized control personnel. Various types of accident plans, monitoring disposal methods and other knowledge in the form of a large number of texts need to be repeatedly memorized and inquired by remote centralized control operation personnel, which is easy to be overlooked or omitted, has low execution efficiency, and the operation experience and ability are difficult to share and inherit, so that the accuracy and standardization of accident handling are difficult to guarantee. In terms of safety prevention, under the remote centralized control mode, it is inevitable to give orders or remotely operate, especially in emergency situations such as accident handling, it is particularly necessary to strengthen the prevention of errors, otherwise there is a risk of safety checking and operation. In terms of flood dispatching, there is no systematic dispatching knowledge base, and flood dispatching may have problems such as untimely response and insufficiently refined dispatching strategy.

[0003] The abnormal information emergency disposal, operation error prevention and flood dispatching in the work of remote centralized control of hydropower stations involve many types of heterogeneous data, including device topology structure, device operation state, reservoir inflow condition, emergency disposal plan, dispatching regulation details and the like. Different data sources, different formats or different structures of data bring difficulties to the intelligent improvement of remote centralized control of hydropower stations. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a method for constructing a knowledge graph of centralized control and dispatching services based on machine learning, to structurally represent various types of faults, abnormalities and early warning information, and to establish a knowledge graph including device fault types, diagnosis rules, disposal schemes and the like, so as to help the centralized control and dispatching system to quickly respond and efficiently dispose.

[0005] The technical solution of the present application is:

[0006] A method for constructing a knowledge graph of centralized control and dispatching services based on machine learning, the method comprising:

[0007] Step 1, construction of a device entity graph: extracting entity triple data related to devices, and converting the entity triple data into a graph database to realize the construction of a device entity graph;

[0008] Step 2, construction of a concept graph: taking unstructured data of classified abnormal emergency disposal and error prevention rules as input and assisted by the constructed device entity graph, automatically labeling the relationship between entities, and on this basis, completing the construction of a concept graph through weakly supervised machine learning;

[0009] Step 3, business logic graph construction: through machine learning to mine the association information of operation ontology and state ontology, and further integrate device entity graph to complete the construction;

[0010] Step 4, case graph construction: using machine learning-based event extraction method to represent training samples through feature vectors, training classification model through supervised learning method, and extracting event instances from unlabeled text to complete the construction of case graph.

[0011] The construction of device entity graph in step 1 includes:

[0012] Step 1.1, determine the field and target: the field is device management and operation, and the target is to construct device entity graph;

[0013] Step 1.2, collect and organize data related to devices: including text data and structured data. Text data includes device manuals, device repair manuals and device operation manuals; structured data includes voltage, power and frequency attribute information of devices;

[0014] Step 1.3, design ontology and schema: design ontology and schema to describe device entities and their relationships; ontology is used to describe the concepts and attributes of entities in the field, and schema is used to describe the relationships between ontologies;

[0015] Step 1.4, entity recognition and attribute extraction: entity recognition and attribute extraction are performed on text data, entity recognition refers to extracting entity names and types from text data, and attribute extraction refers to extracting attribute information of entities from text data;

[0016] Step 1.5, relationship extraction and generation of triples: through neural network extraction method based on device text features, relationship information between entities is extracted from text data, and these relationship information is converted into triples;

[0017] Step 1.6, graph storage and query: store the triples of entities and relationships into graph database;

[0018] Step 1.7, the attributes and connection relationships of devices are updated according to real-time operation data, and the device entity graph is updated after using real-time data for updating.

[0019] Concept graph construction includes:

[0020] Step 2.1, relationship extraction based on distance supervised learning: including:

[0021] Relationship annotation: the relationship between known entity pairs is annotated as positive examples, and the relationship between unknown entity pairs is annotated as negative examples;

[0022] Feature extraction: extract information related to the relationship between entity pairs in the text for training machine learning models;

[0023] Model training: use the annotated entity pair data and corresponding features to train a machine learning model; divide the data into training set and test set when training the model, use the training set to train the model, and use the test set to evaluate the performance of the model; the model used is a random forest model;

[0024] Relation extraction: compare the extracted entity pairs with the distance of known entity pairs to obtain the relationship between entities;

[0025] Step 2.2, after extracting entities and relationships, construct a concept map based on the extracted entities and relationships; when constructing the concept map, a weakly supervised machine learning method is used, a part of the data that has been labeled is used as the training set, the relationship extraction model is used to predict the relationship of the unlabeled data, and the predicted result is compared with the known label, so as to adjust the parameters and prediction results of the model through the feedback mechanism, and finally obtain an accurate concept map.

[0026] Business logic map construction includes:

[0027] Step 3.1, machine learning classification based on text semantics; perform word segmentation and preprocessing on the text semantics to be processed, then convert the preprocessed text into vector representation, and finally use machine learning algorithm to classify and extract the text; when classifying, input the feature vector into the machine learning classification model for classification and extraction; support vector machine SVM is used for classification;

[0028] Step 3.2, knowledge discovery algorithm based on neural tensor network model, through machine learning mining, the association information of the ontology and the state ontology is obtained, and the device entity graph is fused to complete the construction.

[0029] The training process of support vector machine SVM is realized by optimizing the objective function, and the method of quadratic programming is used to solve.

[0030] The construction of case map includes:

[0031] Step 4.1, event extraction based on machine learning: automatically identify and extract event instances from unstructured text data and convert them into structured form;

[0032] Step 4.2, after extracting the time instance, construct the case map according to the extracted event instance data.

[0033] The event extraction based on machine learning includes:

[0034] Feature extraction: Represent the text as a feature vector, using the bag-of-words model, each word as a feature, and the number of occurrences as the weight of the feature.

[0035] Training the classification model: Use support vector machines to train the model.

[0036] Extracting event instances: Apply the trained model to unlabeled text data to extract event instances, the model output includes event type, trigger word, related entity and entity role information.

[0037] The method for constructing a case graph includes:

[0038] Step 4.2.1, define the case representation: define a structured case representation to describe abnormal information, dispatching operation and related entities and attributes; the case representation includes event type, trigger word, entity and entity role, and a case is represented as a tuple:

[0039]

[0040] In the formula, E represents the event type, T represents the trigger word, A represents the set of related entities, R represents the relationship between entities, and the association between entities and events;

[0041] Step 4.2.2, extract case features: for each extracted event instance, extract features and represent them as structured cases, the extracted features include lexical features, syntactic features and semantic features;

[0042] Step 4.2.3, store the structured case representation into the database as a case base;

[0043] Step 4.2.4, calculate the case similarity: define a case similarity, use cosine similarity, Euclidean distance or Manhattan distance to measure the similarity;

[0044] Step 4.2.5, case graph fast matching: when a new abnormality or operation process occurs, extract the features of the case at the concept level and calculate the case similarity for matching, and push the disposal history and operation suggestion of similar cases.

[0045] The beneficial effects of the present application are:

[0046] According to the characteristics of water and electricity remote centralized control and dispatching, the present application constructs entity graph, concept graph, business logic graph and case graph; the concepts, entities, events and their relationships in centralized control and dispatching business are described in a structured way, which provides a knowledge base for intelligent emergency disposal of abnormal events and intelligent safety and error prevention process of operation.

[0047] The knowledge graph constructed by the application has great advantages in solving the safety problems existing in the remote centralized control of hydropower. Through machine learning analysis of water and electricity field related data and accurate analysis of the relationship of each entity in the data, the construction difficulty of the knowledge graph is effectively reduced and the accuracy is improved, which provides protection for the emergency disposal, flood regulation and other behaviors of the remote centralized control of hydropower. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 A flowchart for constructing the application is provided.

[0049] Figure 2 A flowchart for constructing the mode layer and the data layer of the application is provided. DETAILED DESCRIPTION

[0050] The application provides a centralized control scheduling business knowledge graph construction method based on machine learning, including the construction of device entity graph, concept graph, business logic graph and case graph.

[0051] The data for constructing the graph of the application includes: data preprocessing, the first step, data cleaning, eliminating noise and inconsistency in the data. This step includes processing missing values, outliers or duplicate records to ensure the accuracy and reliability of the data. Missing values are usually completed by interpolation method, while outliers need to be identified and decided whether to delete or smooth. For duplicate records, removal is required. The cleaned structured numerical data is converted into text format. Structured data integration is achieved by constructing descriptive sentences, which contain timestamps, data points and other related information that may affect the data. After preprocessing, sequence data and text data are used to mine data features and rules.

[0052] Step 1, construction of device entity graph:

[0053] The device entity graph mainly includes primary or secondary device entities related to power generation, the connection relationship between each entity, and the attributes of each entity. At the same time, the attributes and connection relationship of the device such as voltage, power and frequency can be updated according to real-time operation data. A top-down knowledge graph construction method is adopted, and a neural network extraction method based on device text features is used to automatically represent and capture the effective features of device entities from structured and unstructured data, extract entity triple data related to devices, and convert these entity triple data into a graph database to realize the construction of device entity graph. The specific implementation includes:

[0054] Step 1.1, determine the field and target;

[0055] The field and target of the knowledge graph to be constructed are determined. In the application, the field is device management and operation, and the target is to construct a device entity graph.

[0056] Step 1.2, Collect and organize data;

[0057] In order to construct the device entity graph, it is necessary to collect and organize data related to the device, including text data and structured data. Text data can be device manuals, device maintenance manuals, device operation manuals, etc., while structured data can be device voltage, power, frequency, etc. attribute information.

[0058] The device entity knowledge graph first needs to comprehensively integrate the data of the hydropower station equipment. Including collecting data from various sensors, operation logs and historical maintenance records, and extracting text and structured data.

[0059] Step 1.3, design ontology and schema;

[0060] After determining the data source, the ontology and schema need to be designed to describe the device entities and their relationships. Ontology is used to describe the concepts and attributes of entities in the field, while schema is used to describe the relationships between ontologies.

[0061] Step 1.4, entity recognition and attribute extraction;

[0062] Using natural language processing technology, entity recognition and attribute extraction are performed on text data. Entity recognition refers to extracting the name and type of an entity from text data, such as "generator", "transformer", etc. Attribute extraction refers to extracting attribute information of an entity from text data, such as voltage, power, frequency, etc.

[0063] Step 1.5, relationship extraction and generation of triples;

[0064] According to the schema, the invention extracts relationship information between entities from text data through a neural network extraction method based on device text features, and converts these relationship information into triples.

[0065] The neural network extraction based on device text features adopted by the invention uses a deep learning model to perform entity and relationship extraction. The relationship extraction model maps entities and relationships in the text to a high-dimensional space, and then calculates the similarity between entities and relationships to obtain the relationship between entities.

[0066] After completing the extraction of entities and relationships, triples are generated. The generation method includes: first, aligning the entities in the relationship with the known entities in the device entity graph to determine the unique identifier of the entity. Then determine the relationship type, such as "connected to", "belongs to", etc. Finally, according to the entity alignment and relationship type, convert the relationship into triples. For example, for the relationship "generator connected to transformer", triples (generator, connected to, transformer) can be generated.

[0067] Step 1.6, graph storage and query;

[0068] The triple data of entities and relationships is stored into a graph database, such as Neo4j. Then the information in the device entity graph can be queried using the query interface provided by the graph database.

[0069] Step 1.7, real-time data update;

[0070] The attributes and connection relationships of the device can be updated according to the real-time running data. After using the real-time data for updating, the device entity graph needs to be updated to maintain the accuracy and real-time of the data.

[0071] The device entity graph constructed by the present application is a key graph that describes the connection mode of the device. Based on this graph, various functions such as intelligent troubleshooting of device historical data, simulation of fault scenarios, and detection of potential dangerous points in the power station circuit can be realized.

[0072] The device entity knowledge graph uses nodes to maximize the replication of each device in the power grid, and stores real-time key parameters and historical parameter sequences of device operation in the node attributes. For example, in the "hydro-generator" node, the model, speed, rated speed, and other dynamic and static data of the generator are stored. Other modules can conveniently and quickly call the information in the device entity graph, update the node attributes after sufficient data mining, and realize intelligent knowledge update and knowledge reasoning.

[0073] Step 2, concept graph construction:

[0074] The concept graph of the present application is an abstraction of the entity graph, which is used to standardize and refine a series of fact expressions in the data layer. The rules, logic, and case judgments constructed need to be operated in the concept layer first, and then mapped to the corresponding entity graph for specific information matching. A bottom-up knowledge graph construction method is adopted, a relationship extraction method based on distance supervised learning is adopted, unstructured data of classification exception emergency disposal and anti-misoperation rules is taken as input and supplemented by the constructed device entity graph, automatic annotation of the relationship between entities is performed, and on this basis, weakly supervised machine learning is completed to construct the concept graph, which specifically includes:

[0075] Step 2.1, relationship extraction based on distance supervised learning, including:

[0076] Relationship annotation: the relationship between known entity pairs is annotated as positive examples, and the relationship between unknown entity pairs is annotated as negative examples. The purpose of relationship annotation is to train a machine learning model to identify the relationship between entities. Relationship annotation usually needs to use domain knowledge or external knowledge base, such as WordNet, Freebase, etc.

[0077] Feature extraction: Extract information related to the relationship between entity pairs in the text for training machine learning models. Features include the distance between entities, the frequency of entity appearance in the text, the similarity of entities in the vocabulary, etc.

[0078] Model training: Use annotated entity pair data and corresponding features to train a machine learning model. When training the model, the data needs to be divided into training set and test set, using the training set to train the model, and using the test set to evaluate the performance of the model. For the concept graph of the present invention, the model used is a random forest model. By randomly sampling the data and randomly selecting features, multiple decision trees are constructed, and the results of multiple decision trees are finally integrated, thereby improving the robustness and generalization ability of the model.

[0079] Relation extraction: Compare the distance between the extracted entity pairs and the known entity pairs to determine the relationship between the entities. By setting a threshold, if the distance is less than a certain threshold, it is considered that there is a relationship between the entities.

[0080] Step 2.2, construction of concept graph;

[0081] After the extraction of entities and relationships, the construction of the concept graph is carried out according to the extracted entities and relationships.

[0082] The concept graph is an abstraction of the entity graph, used to standardize and refine a series of fact expressions in the data layer. The rules, logic and case judgment need to be operated in the concept layer first, and then mapped to the corresponding entity graph for specific information matching.

[0083] In the construction of the concept graph, a weakly supervised machine learning method is used. A part of the data that has been labeled is used as the training set, and the relationship extraction model is used to predict the relationship of the unlabeled data, and the predicted results are compared with the known labels, so as to continuously adjust the parameters and prediction results of the model through the feedback mechanism, and finally obtain an accurate concept graph.

[0084] The water regulation concept graph and the electric regulation concept graph are constructed by using the above construction method,

[0085] The water regulation concept graph is a machine language formed by abstracting the water regulation rules and the logic of the cascade hydropower station, forming a flood regulation concept graph. Compared with the business logic graph, it can flexibly represent various entities such as power stations, water conditions and regulation rules, and has the ability of automatic updating and expansion, guiding the generation of business logic graph.

[0086] The electric regulation concept map: the concept center node representing the hydropower station is extended outward, and various concepts are preliminarily classified and stored in the concept classification node, and then specific concept nodes are extended downward. The concept center node is the entity corresponding to the concept, and the concept center node stores the relevant important parameters and retrieval identifiers of the corresponding entity; the concept classification node mainly stores the retrieval identifier, so as to quickly locate the related concept; the specific concept node stores the concept pointing entity, judgment index and the like.

[0087] The concept map obtained by the application can concretize abstract concepts, clarify relationships and enrich attributes, and enhance the understanding and control of the complexity of the hydropower control, and improve the scientificity and efficiency of decision-making.

[0088] Step 3, business logic map construction:

[0089] The business logic map is knowledge extracted from existing dispatching specifications, dispatching rules, switching operation regulations, emergency disposal plans and monitoring disposal manuals and the like, and when an abnormal situation occurs, related information query and reasoning are performed in the knowledge path of the map network. Based on the text semantic machine learning classification extraction method, the knowledge discovery algorithm based on the neural tensor network model is used to mine the associated information of the operation ontology and the state ontology by machine learning, and further fuse the device entity map to complete the construction; specifically including:

[0090] Step 3.1, machine learning classification extraction based on text semantics;

[0091] First, the text semantics to be processed are segmented and preprocessed, then the preprocessed text is converted into a vector representation, and finally the machine learning algorithm is used to classify and extract the text. In the classification, the feature vector is input into the machine learning classification model for classification and extraction.

[0092] The application adopts support vector machine (SVM) for classification.

[0093] The SVM of the application introduces a kernel function, which avoids direct calculation in a high-dimensional space. Common kernel functions include linear kernel, polynomial kernel, radial basis kernel and the like, and their respective formulas are as follows:

[0094] ;

[0095] ;

[0096] ;

[0097] Among them, and denote two vectors in the input space, and are parameters of the polynomial kernel, is the parameter of the radial basis kernel.

[0098] The training process of SVM can be achieved by optimizing the objective function, which is generally solved using the method of quadratic programming. After obtaining the optimal classification hyperplane, it can be used to classify new input data.

[0099] Step 3.2, the knowledge discovery algorithm based on Neural Tensor Network (NTN), mines the associated information of the ontology and state ontology through machine learning, and further integrates the device entity graph to complete the construction.

[0100] The input of NTN model contains two entity vectors and a relation vector, which are combined through a tensor product operation to obtain a scalar value representing the strength of the relationship between the two entities. Specifically, the input of NTN model consists of the following three parts:

[0101] (a) Entity vectors and : represent the feature vectors of two entities, with dimension .

[0102] (b) Relation vector : represents the feature vector of the relationship, with dimension .

[0103] (c) Matrix : used to capture the interaction between the relation vector and the entity vector.

[0104] After inputting through the tensor product operation, a scalar value can be obtained, representing the strength of the relationship between entities and :

[0105] ;

[0106] where denotes concatenating two vectors, denotes the matrix, denotes the nonlinear activation function, usually the hyperbolic tangent function or the ReLU function.

[0107] Finally, the output of NTN model is a dimensional vector, where represents the number of relationship categories. The output can be converted into a probability distribution through the softmax function for classification:

[0108] ;

[0109] The training of the NTN model can adopt a back propagation algorithm to optimize the model parameters by minimizing a cross-entropy loss function. Specifically, for a training sample consisting of

[0110] ;

[0111] wherein, denotes the training set, denotes whether the entity and has a relationship.

[0112] The trained neural tensor network model is used to predict the association between the operation ontology and the state ontology, and to mine the key features therein. The association information between the operation ontology and the state ontology is found through these features, and a logical map of the operation ontology and the state ontology is constructed.

[0113] The application constructs an electric regulation service logic knowledge graph and a water regulation service logic knowledge graph.

[0114] The storage nodes of the electric regulation service logic knowledge graph can be divided into four categories of preplan, situation, phenomenon and operation, and different nodes are marked by a label method.

[0115] The preplan node is relatively short and only stores the general information of the preplan, only contains name attribute and description attribute, name is used to store the preplan name, and description represents details.

[0116] The situation node is the situation corresponding to a set of operations, since a preplan can contain multiple situations, each situation has a corresponding solution method, and the situation node is defined separately. The situation node contains name attribute and description attribute, the name attribute is used for situation name, and the description represents details.

[0117] The phenomenon node is the starting condition of each case. When the starting condition is met in whole or in part, it is considered that the case corresponding to the phenomenon node occurs, and then the corresponding solution can be found from the case node, including name, description, task, mon_, and fixed attributes. The M name attribute is used to store the information displayed in the most intuitive position; description indicates details; task is a list data used to store the sub-phenomenon of the phenomenon to improve the accuracy of phenomenon judgment; mon_ is a series of variable attributes, all of which start with mon_, such as mon_repair_pump and mon_dam_leakage_pum. These attributes are used to store the monitoring data of the related equipment; and fixed is the alarm standard quota corresponding to the equipment variable attribute, such as the rated speed of the water turbine, which is convenient for direct extraction and display.

[0118] Each operation node includes an operation step and related keyword labels. Among these keyword labels, the most important one is the label referring to the operation subject, such as "dispatching officer", "power station side", etc.

[0119] The operation node includes name, description, task, time_limit, and contact attributes; name, description, and task have the same meaning as the phenomenon node; time_limit is used to indicate whether there is a definite time limit for the step, with the unit being minutes, and -1 when not defined; and contact indicates whether there is a related contact for the operation, and an empty list when not defined.

[0120] Compared with the power dispatching knowledge graph, the water dispatching knowledge graph has differences in structure and function, and the focus is on returning all operations involved in the regulation under specific environmental conditions.

[0121] The water dispatching knowledge graph is constructed in consideration of improving search efficiency and saving storage space, and the construction form includes:

[0122] (1) In flood dispatching, hydrological element entities such as inflow and basin rainfall complement each other and play an important role. This complementary nature makes the node reuse degree of the water dispatching knowledge graph higher, and the structure is in a network form rather than a traditional tree structure.

[0123] (2) Unlike the binary judgment of "yes" or "no" in the power dispatching knowledge graph, the key condition judgment in flood dispatching is often based on interval distribution in a ladder shape, resulting in a large number of conditions. Based on the above characteristics of the logic, the conditions triggered by the flood dispatching mechanism are not displayed as independent nodes in the graph, but are stored in the form of attributes in the relationship.

[0124] (3) Water regulation knowledge graph is divided into basin, and the knowledge graph is divided into three parts of Beipanjiang basin, Sanfah river basin and Furongjiang river basin. Each basin takes the alarm level and operation of the basin as the core, and builds knowledge context around each hydropower station. The inflow is an important factor for determining the reservoir regulation, and a node is created to store the inflow data. "Edge" is used to store specific inflow and rainfall determination range.

[0125] Step 4, construction of case graph:

[0126] The case graph records and saves each abnormal information and operation, and forms a structured case library. The event extraction method based on machine learning is studied. The training samples are represented by feature vectors. The classification model is trained by supervised learning method. The event instances are extracted from the unmarked text to complete the construction of the case graph. When a new abnormality or operation process occurs, the features of the case are extracted in the concept layer and the case similarity is calculated, so as to push the disposal history and operation suggestion of similar cases. Specifically, it includes:

[0127] Step 4.1, event extraction based on machine learning: automatically identify and extract event instances from unstructured text data, and convert them into structured form for storage and analysis. Specifically, it includes:

[0128] Feature extraction: represent text as feature vector, which can use bag-of-words model, where each word is a feature and the number of occurrences is the weight of the feature. Suppose there are texts and words, a feature matrix can be constructed, where represents the number of occurrences of the word in the text.

[0129] Training classification model: use support vector machine to train the model;

[0130] Extract event instances: apply the trained model to unmarked text data to extract event instances. The model output includes event type, trigger word, related entity and entity role, etc. These information can be used to build case graph and push disposal history and operation suggestion of similar cases.

[0131] Step 4.2, extract events to build case graph; after extracting the event instance, the case graph is constructed according to the extracted event instance data. The steps of constructing case graph mainly include:

[0132] (1) Define case representation: define a structured case representation for describing abnormal information, scheduling operations and their related entities and attributes. Case representation includes event type, trigger word, entity and entity role, etc. For example, a case is represented as a tuple:

[0133]

[0134] In the formula, E represents the event type (such as device failure and system exception, etc.), T represents the trigger word (such as "failure" and "alarm", etc.), A represents the set of related entities (such as devices, components and sensors, etc.), R represents the relationship between entities and the association between entities and events (such as "failure cause" and "impact range", etc.).

[0135] (2) Extract case features:

[0136] For each extracted event instance, extract features and represent them as structured cases. The extracted features include lexical features, syntactic features and semantic features, etc. Using these features helps to calculate the similarity between cases.

[0137] (3) Build case library

[0138] Store the structured representation of the case into the database as a case library for storing and retrieving cases.

[0139] (4) Calculate case similarity: the invention defines a case similarity, which uses cosine similarity, Euclidean distance, or Manhattan distance measurement. According to the nature of the features and task requirements, select the appropriate measurement method. Taking cosine similarity as an example, given two cases C1 and C2, the feature vectors are V1 and V2, and the cosine similarity calculation formula is:

[0140] ;

[0141] (5) Case atlas rapid matching:

[0142] When a new abnormality or operation process occurs, the case is extracted in the concept layer and the case similarity is calculated for matching, and the disposal history and operation suggestion of similar cases are pushed.

[0143] The method of case atlas rapid matching includes:

[0144] New case feature extraction:

[0145] When a new abnormality or operation process occurs, first extract the features of the new case. The extracted features include lexical features, syntactic features and semantic features, etc., which are consistent with the features used when building the case library.

[0146] (b) Similarity calculation: calculate the similarity between each case in the case library and the new case , . The similarity between the index of the case and the new case is calculated. Cosine similarity is used as the similarity measure, given two cases and , the feature vectors are and , and the cosine similarity calculation formula is:

[0147] ;

[0148] (c) Set the similarity threshold:

[0149] In order to ensure that the recommended case has high similarity, a similarity threshold τ is set. Only when the similarity between a case in the case library and the new case is greater than the threshold, it is recommended as a recommended case. The threshold is set according to actual demand and experience.

[0150] (d) Sorting and recommendation:

[0151] According to the similarity, the cases in the case library are sorted. Select the case whose similarity is higher than the threshold τ for recommendation. Multiple similar cases can be recommended to provide more reference information and operation suggestions for the user.

[0152] The construction of the device entity graph, the concept graph, the business logic graph and the case graph of the present application includes mode layer construction and data layer construction. The mode layer is an organizational framework for describing entity concepts, attributes and relationships between entities. The data layer construction is to select appropriate methods to extract entities and relationships from emergency knowledge. The entities and relationships in the data layer are associated and mapped according to the preset mode of the mode layer, so as to obtain the knowledge graph. The mode layer and the data layer are constructed in a top-down and bottom-up manner as shown in Figure 1 .

[0153] Mode layer construction:

[0154] The present application takes the business logic graph as an example, takes the four stages of emergency disposal process in centralized control and dispatching business as the main clue, analyzes the entity concepts and attributes concerned in each stage, constructs the business logic knowledge graph mode layer, extracts the corresponding entities and attributes from the emergency plan in each stage of emergency disposal, and thus realizes the construction of the emergency disposal knowledge graph.

[0155] Data layer construction: The centralized scheduling business knowledge graph can be expressed by G=(E, R, S), wherein E is an entity set in the knowledge base, R is a relation set, and S is a "node-relation-node" triple knowledge set.Data layer construction is to extract the required entities and relations from the related text under the guidance of the mode layer organization framework.

[0156] The application finally displays the visual knowledge graph by using a graph database Neo4j, can store the extracted entities and relations, and visually represents the "node-relation-node" triple knowledge, has the characteristics of good visual effect and efficient search, and can find each node and relation through a query statement, helps decision makers to quickly determine the work to be performed and the responsible department at each stage of disposal, and greatly improves the efficiency.

Claims

1. A method for constructing a centralized control and scheduling business knowledge graph based on machine learning, characterized in that: The method includes: Step 1: Construction of Equipment Entity Map: Extract the entity triple data related to the equipment and transform this entity triple data into a graph database to realize the construction of the equipment entity map; Step 2, Concept Graph Construction: Using unstructured data on classification anomaly emergency response and error prevention rules as input, and supplemented by the constructed equipment entity graph, the relationships between entities are automatically labeled, and weakly supervised machine learning is performed on this basis to complete the construction of the concept graph; Step 3, Business Logic Graph Construction: Machine learning is used to mine the correlation information between operation ontology and state ontology, and the device entity graph is further integrated to complete the construction; Step 4: Construction of the case graph: The training samples are represented by feature vectors using a machine learning-based event extraction method. A classification model is trained using a supervised learning method. Event instances are then extracted from the unlabeled text to complete the construction of the case graph.

2. The method for constructing a centralized control and scheduling business knowledge graph based on machine learning according to claim 1, characterized in that: Step 1 involves constructing the device entity map, which includes: Step 1.1: Define the domain and objectives: The domain is equipment management and operation, and the objective is to construct an equipment entity map; Step 1.2: Collect and organize equipment-related data, including text data and structured data. Text data includes equipment manuals, equipment maintenance manuals, and equipment operation manuals; structured data includes equipment voltage, power, and frequency attribute information. Step 1.3, Design Ontologies and Schemas: Design ontologies and schemas to describe device entities and the relationships between them; ontologies are used to describe the concepts and attributes of entities in the domain, and schemas are used to describe the relationships between ontologies; Step 1.4, Entity Recognition and Attribute Extraction: Perform entity recognition and attribute extraction on the text data. Entity recognition refers to extracting the name and type of the entity from the text data, and attribute extraction refers to extracting the attribute information of the entity from the text data. Step 1.5, Relation Extraction and Triple Generation: Extract the relation information between entities from the text data using a neural network extraction method based on device text features, and convert this relation information into triples; Step 1.6, Graph Storage and Query: Store the triples of entities and relations in the graph database; Step 1.7: Update the device attributes and connection relationships based on real-time operating data, and then update the device entity map using the updated real-time data.

3. The method for constructing a centralized control and scheduling business knowledge graph based on machine learning according to claim 1, characterized in that: Concept map construction includes: Step 2.1, Relation Extraction Based on Distance-Supervised Learning: Includes: Relationship labeling: Label relationships between known entity pairs as positive examples, and relationships between unknown entity pairs as negative examples; Feature extraction: Extracting information from text that relates to the relationships between entity pairs for use in training machine learning models; Model training: A machine learning model is trained using labeled entity pairs and their corresponding features; during model training, the data is divided into training and test sets, the training set is used to train the model, and the test set is used to evaluate the model's performance; the model used is a random forest model; Relation extraction: The distance between the extracted entity pairs and known entity pairs is compared to determine the relationships between the entities; Step 2.2: After extracting entities and relationships, construct a concept graph based on the extracted entities and relationships. When constructing the concept graph, a weakly supervised machine learning method is adopted. A portion of the already labeled data is used as the training set. The relationship extraction model is used to predict the relationships of the unlabeled data. The prediction results are compared with the known labels. The feedback mechanism is used to continuously adjust the parameters of the model and the prediction results, and finally obtain an accurate concept graph.

4. The method for constructing a centralized control and scheduling business knowledge graph based on machine learning according to claim 1, characterized in that: The construction of the business logic graph includes: Step 3.1: Machine learning classification and extraction based on text semantics; The text to be processed is segmented and preprocessed, then the preprocessed text is converted into a vector representation, and finally the text is classified and extracted using machine learning algorithms; During classification, the feature vectors are input into the machine learning classification model for classification and extraction; Support Vector Machine (SVM) is used for classification; Step 3.2: The knowledge discovery algorithm based on the neural tensor network model mines the correlation information between the operation ontology and the state ontology through machine learning, and integrates the device entity graph to complete the construction.

5. The method for constructing a centralized control and scheduling business knowledge graph based on machine learning according to claim 4, characterized in that: The training process of Support Vector Machine (SVM) is achieved by optimizing the objective function, which is solved using a quadratic programming method.

6. The method for constructing a centralized control and scheduling business knowledge graph based on machine learning according to claim 1, characterized in that: The construction of the case graph includes: Step 4.1, Machine learning-based event extraction: Automatically identify and extract event instances from unstructured text data and transform them into a structured form; Step 4.2: After extracting the time instances, construct a case graph based on the extracted event strength data.

7. The method for constructing a centralized control and scheduling business knowledge graph based on machine learning according to claim 5, characterized in that: The machine learning-based event extraction includes: Feature extraction: The text is represented as a feature vector. The bag-of-words model is used, where each word is treated as a feature and the frequency of its occurrence is used as the weight of that feature. Training the classification model: The model is trained using a support vector machine; Event instance extraction: The trained model is applied to unlabeled text data to extract event instances. The model output includes event type, trigger words, related entities, and entity role information.

8. The method for constructing a centralized control and scheduling business knowledge graph based on machine learning according to claim 5, characterized in that: Methods for constructing case graphs include: Step 4.2.1, Define Case Representation: Define a structured case representation to describe exception information, scheduling operations, and related entities and attributes; the case representation includes event type, trigger word, entity, and entity role, and represents a case as a tuple: ； In the formula, E represents the event type, T represents the trigger word, A represents the set of related entities, and R represents the relationship between entities and the association between entities and events; Step 4.2.2: Extract case features: For each extracted event instance, extract features and represent them as a structured case. The extracted features include lexical features, grammatical features, and semantic features. Step 4.2.3: Store the structured representation of the cases in the database as a case library; Step 4.2.4: Calculate case similarity: Define a case similarity and use cosine similarity, Euclidean distance or Manhattan distance to measure similarity; Step 4.2.5, Case Graph Quick Matching: When a new anomaly or operation occurs, the case will be matched by extracting features at the conceptual level and calculating case similarity. The handling history and operation suggestions of similar cases will be pushed.

Images