CN-122020275-A - Multi-mode AI fusion heat supply intelligent diagnosis method

CN122020275ACN 122020275 ACN122020275 ACN 122020275ACN-122020275-A

Abstract

The invention provides a heat supply intelligent diagnosis method with multi-mode AI fusion, which relates to the technical field of data processing, and comprises the steps of constructing a structured knowledge graph comprising a physical topology layer, an operation characteristic layer, a fault mode layer and a regulation strategy layer according to heat supply operation and maintenance data; based on the structured knowledge graph, a multi-mode AI diagnosis model group is constructed, the structural relation and attribute characteristics of the knowledge graph are embedded into a graph neural network, and a XGBoost algorithm is fused for model training, so that a trained model for heating a plurality of diagnosis scenes is obtained. The invention realizes the normal form transition of heat supply diagnosis from manual decision to digital intelligence driving.

Inventors

LIU YUBIN
YIN XIANZHEN
ZHANG YU
FU QIANG
MEI DEFANG
WU SHENGJUN
YANG GAISHUN
YAN DONGXU
ZHANG JIANHUA
FENG MINGRU
ZHAO NAN

Assignees

北京京能热力股份有限公司

Dates

Publication Date: 20260512
Application Date: 20251216

Claims (10)

1. A multi-modal AI-converged heating intelligent diagnostic method, the method comprising: Constructing a structured knowledge graph comprising a physical topology layer, an operation characteristic layer, a fault mode layer and a regulation strategy layer according to the heat supply operation and maintenance data; constructing a multi-mode AI diagnosis model group based on the structured knowledge graph, embedding the structural relationship and attribute characteristics of the knowledge graph into a graph neural network, and fusing XGBoost algorithm to perform model training to obtain a trained model for heating a plurality of diagnosis scenes; The method comprises the steps of acquiring real-time data of heat supply operation based on a trained model, converting the real-time data into a data unit set with multidimensional features, constructing a feature distribution characterization layer based on the data unit set, dividing the characterization layer into a plurality of feature analysis areas through a space division algorithm, dividing the area into a plurality of feature subspaces according to data features, mapping the data units to the corresponding feature subspaces, generating dynamic optimization parameters according to the data distribution characteristics of each subspace, performing feature optimization on the original real-time data through the dynamic optimization parameters to obtain optimized feature data, inputting the optimized feature data into the trained model for analysis, and obtaining intelligent diagnosis results; Based on the diagnosis accuracy and the artificial check feedback of the intelligent diagnosis result, a bidirectional evaluation mechanism of diagnosis accuracy and knowledge completeness is established to obtain an evaluation result, and dynamic iterative optimization is carried out on the multi-mode AI diagnosis model group and the structured knowledge graph simultaneously according to the evaluation result to obtain an improved intelligent diagnosis closed loop.
2. The multi-modal AI-converged heating intelligent diagnosis method of claim 1, wherein constructing a structured knowledge graph including a physical topology layer, an operation feature layer, a failure mode layer, and a regulation policy layer according to heating operation and maintenance data comprises: The physical topology layer is constructed by extracting all physical entities including a heat source, a pipe network, a heat exchange station and an end user based on heat supply network design data and equipment list, defining subordinate and connection relations among the entities, obtaining a topology network comprising all physical entities and connection relations of the heat source, the pipe network, the heat exchange station and the end user, and taking the topology network as a structural basis of a knowledge graph; the construction of the operation feature layer is based on historical operation data and equipment parameters, and configures corresponding static attribute and dynamic operation parameters for each key physical entity in the physical topology layer; The fault mode layer is constructed by defining a known fault mode based on heat supply operation data and a historical fault record, associating each fault mode with one or more abnormal characteristic states with causal relation in the operation characteristic layer, and constructing a fault mode and characteristic association library; The method comprises the steps of searching and matching corresponding regulation strategy knowledge entities based on various fault modes in a fault mode and feature association library, constructing an inference path from fault diagnosis to strategy generation based on the regulation strategy knowledge entities through semantic association binding, and integrating and fusing the inference path with a physical topology layer, an operation feature layer and a fault mode layer to form a structured knowledge map with closed-loop diagnosis and regulation capability.
3. The multi-modal AI-converged heating intelligent diagnosis method of claim 2, wherein the multi-modal AI diagnosis model group is constructed based on the structured knowledge graph, the structural relationship and the attribute characteristics of the knowledge graph are embedded into the graph neural network, and model training is performed by fusing XGBoost algorithm, so as to obtain a trained model for heating a plurality of diagnosis scenes, comprising: Based on a physical topology layer of the structured knowledge graph, extracting connection relations among equipment entities in a pipe network, constructing a graph structure of a graph neural network, and forming graph data representing a hydraulic transmission path by taking the equipment entities as nodes and the connection relations as edges; Configuring corresponding operation characteristic attributes for nodes in the graph based on the graph data and an operation characteristic layer of the structured knowledge graph, and obtaining a characteristic enhancement rule set cooperated with the graph structure based on the heat supply operation and maintenance data; Injecting the feature enhancement rule set into feature engineering of XGBoost algorithm, and optimizing the operation feature attribute to obtain enhancement feature set with physical interpretability; and carrying out multi-scene model training by fusing the map neural network and XGBoost algorithm of the injection enhanced feature set through the map data and the enhanced feature set to obtain a trained model for work order classification, hydraulic balance diagnosis and heat source load prediction.
4. The multi-modal AI-fusion heating intelligent diagnosis method of claim 3, wherein acquiring real-time data of a heating operation based on a trained model, converting the real-time data into a set of data units having multi-dimensional features, constructing a feature distribution characterization layer based on the set of data units, dividing the characterization layer into a plurality of feature analysis regions by a spatial division algorithm, dividing the region into a plurality of feature subspaces according to the data features, and mapping the data units to the corresponding feature subspaces, comprises: Based on the trained model, collecting a real-time data stream of heat supply operation, and converting the real-time data stream into a multi-dimensional characteristic data unit set with a space-time association relation; Based on the multidimensional feature vector set, calculating the mean, variance and skewness statistical moment of the multidimensional feature vector set, and constructing a feature distribution state space for representing the global running state; Aiming at each feature vector in the state space, calculating Euclidean distance between each feature vector and the feature center of each typical operation condition, carrying out region attribution division based on a nearest neighbor principle, and dividing the feature distribution state space into a plurality of feature analysis regions corresponding to different operation conditions; And aiming at each characteristic analysis region, analyzing the local density and the data distribution discrete degree of the internal data points, and carrying out secondary division on each region based on a clustering algorithm of density peak detection to obtain a plurality of characteristic subspaces for fine characteristic analysis.
5. The multi-mode AI fusion heating intelligent diagnosis method of claim 4, wherein generating dynamic optimization parameters according to the data distribution characteristics of each subspace, performing feature optimization on the original real-time data through the dynamic optimization parameters to obtain optimized feature data, inputting the optimized feature data into a trained model for analysis and processing to obtain an intelligent diagnosis result, comprising: Based on the distribution characteristics of the data units in the feature subspaces, analyzing the statistical features of each feature subspace, and generating a corresponding dynamic optimization parameter set; according to the dynamic optimization parameter set, performing self-adaptive feature optimization processing on the original real-time data stream to obtain optimized feature data; and inputting the optimized characteristic data into a trained model for analysis and processing to obtain an intelligent diagnosis result comprising the fault type, the fault position and the fault probability.
6. The multi-modal AI-converged heating intelligent diagnosis method of claim 5, wherein establishing a bidirectional evaluation mechanism of diagnosis accuracy and knowledge completeness based on diagnosis accuracy and manual verification feedback of the intelligent diagnosis result to obtain an evaluation result includes: Receiving an intelligent diagnosis result, collecting corresponding manual verification feedback data, and comparing and analyzing the intelligent diagnosis result and the manual verification feedback data to obtain a diagnosis accuracy evaluation index; Based on the diagnosis accuracy evaluation index, performing diagnosis accuracy evaluation, analyzing performance of the multi-mode AI diagnosis model group under each diagnosis scene, and identifying a specific scene and a characteristic mode with insufficient diagnosis accuracy to obtain a diagnosis accuracy evaluation result; Based on the identification result, carrying out knowledge completeness assessment in parallel, and verifying whether a structured knowledge graph contains a fault mode, characteristic association and regulation strategy corresponding to the identified diagnosis precision insufficient scene or not to obtain a knowledge completeness assessment result; And synthesizing the diagnosis precision evaluation result and the knowledge completeness evaluation result to generate a comprehensive evaluation result containing model optimization requirements and knowledge supplement requirements.
7. The multi-modal AI-fusion heating intelligent diagnosis method of claim 6, wherein the dynamic iterative optimization is performed on the multi-modal AI diagnosis model group and the structured knowledge graph simultaneously according to the evaluation result to obtain an improved intelligent diagnosis closed loop, comprising: Based on model optimization requirements in the comprehensive evaluation result, extracting model parameters to be optimized and training data to be supplemented, and carrying out parameter optimization and training data updating on the multi-mode AI diagnosis model group to obtain an optimized multi-mode AI diagnosis model group; based on knowledge supplement requirements in the comprehensive evaluation result, extracting fault mode nodes, characteristic association relations and regulation strategies to be supplemented, and dynamically expanding the nodes and the relations of the structured knowledge graph to obtain an updated structured knowledge graph; And integrating the optimized multi-mode AI diagnosis model group with the updated structured knowledge graph to construct a continuously improved intelligent diagnosis closed loop.
8. A multi-modal AI-converged heating intelligent diagnostic system implementing the method of any one of claims 1 to 7, comprising: the acquisition module is used for constructing a structured knowledge graph comprising a physical topology layer, an operation characteristic layer, a fault mode layer and a regulation strategy layer according to the heat supply operation and maintenance data; The construction module is used for constructing a multi-mode AI diagnosis model group based on the structured knowledge graph, embedding the structural relationship and attribute characteristics of the knowledge graph into the graph neural network, and fusing XGBoost algorithm to perform model training to obtain a trained model for heating a plurality of diagnosis scenes; The system comprises a training module, a diagnosis module, a characteristic distribution characterization layer, a dynamic optimization parameter generation module, a characteristic analysis module and an intelligent diagnosis module, wherein the training module is used for training a model, acquiring real-time data of heat supply operation, converting the real-time data into a data unit set with multidimensional characteristics, constructing the characteristic distribution characterization layer based on the data unit set, dividing the characterization layer into a plurality of characteristic analysis areas through a space division algorithm, dividing the area into a plurality of characteristic subspaces according to data characteristics, mapping the data units to the corresponding characteristic subspaces, generating dynamic optimization parameters according to the data distribution characteristics of each subspace, performing characteristic optimization on original real-time data through the dynamic optimization parameters to obtain optimized characteristic data, and inputting the optimized characteristic data into the training model for analysis processing to obtain an intelligent diagnosis result; The processing module is used for establishing a bidirectional evaluation mechanism of diagnosis precision and knowledge completeness based on diagnosis accuracy and manual check feedback of the intelligent diagnosis result to obtain an evaluation result, and carrying out dynamic iterative optimization on the multi-mode AI diagnosis model group and the structured knowledge graph simultaneously according to the evaluation result to obtain an improved intelligent diagnosis closed loop.
9. A computing device, comprising: one or more processors; Storage means for storing one or more programs which when executed by the one or more processors cause the one or more processors to implement the method of any of claims 1 to 7.
10. A computer readable storage medium, characterized in that the computer readable storage medium has stored therein a program which, when executed by a processor, implements the method according to any of claims 1 to 7.

Description

Multi-mode AI fusion heat supply intelligent diagnosis method Technical Field The invention relates to the technical field of data processing, in particular to a heat supply intelligent diagnosis method with multi-mode AI fusion. Background The traditional diagnosis method relies on manual experience for a long time, more than 1200 high-value diagnosis rules dispersed in operation and maintenance logs and expert interviews of over 20 years are difficult to convert into structured assets, so that experience differences of different operation and maintenance personnel are large, fault judgment consistency is low, meanwhile, a part of pure data driving diagnosis model has difficult problems of physical interpretability due to the fact that the diagnosis model is separated from physical laws such as physical topology and operation characteristics, and a diagnosis strategy cannot evolve in real time along with environmental changes. Taking a certain area of a winter heat supply period of a certain city in the north as an example, the area is low in part of user room temperature and too high in part of user room temperature due to hydraulic unbalance of a building riser, the traditional diagnosis method needs to measure room temperature manually by each user and check the connection relation of pipe networks on site, the unbalanced riser is positioned within 3 days, the standard deviation of the room temperature difference in the building after regulation still reaches 1.8 ℃ and exceeds the national standard of 1.5 ℃, a plurality of user complaint worksheets are generated during regulation, the worksheet processing needs to manually and repeatedly check the user position and the pipe network information, the false dispatch rate is high, the traditional method depends on manual on site operation, the fault positioning efficiency is low and time consuming, the integration of a heat supply physical topological layer and an operation characteristic layer is lacked, the regulation precision is insufficient and is difficult to meet the national standard requirement, the diagnosis and worksheet processing do not form a data closed loop, the real-time feedback optimization strategy cannot be based, the labor cost is high, and the user experience is poor. Disclosure of Invention The technical problem to be solved by the invention is to provide a multi-mode AI fusion heat supply intelligent diagnosis method for realizing the normal form transition of heat supply diagnosis from manual decision to digital intelligent driving. In order to solve the technical problems, the technical scheme of the invention is as follows: In a first aspect, a multi-modal AI-converged heating intelligent diagnostic method includes: Constructing a structured knowledge graph comprising a physical topology layer, an operation characteristic layer, a fault mode layer and a regulation strategy layer according to the heat supply operation and maintenance data; constructing a multi-mode AI diagnosis model group based on the structured knowledge graph, embedding the structural relationship and attribute characteristics of the knowledge graph into a graph neural network, and fusing XGBoost algorithm to perform model training to obtain a trained model for heating a plurality of diagnosis scenes; The method comprises the steps of acquiring real-time data of heat supply operation based on a trained model, converting the real-time data into a data unit set with multidimensional features, constructing a feature distribution characterization layer based on the data unit set, dividing the characterization layer into a plurality of feature analysis areas through a space division algorithm, dividing the area into a plurality of feature subspaces according to data features, mapping the data units to the corresponding feature subspaces, generating dynamic optimization parameters according to the data distribution characteristics of each subspace, performing feature optimization on the original real-time data through the dynamic optimization parameters to obtain optimized feature data, inputting the optimized feature data into the trained model for analysis, and obtaining intelligent diagnosis results; Based on the diagnosis accuracy and the artificial check feedback of the intelligent diagnosis result, a bidirectional evaluation mechanism of diagnosis accuracy and knowledge completeness is established to obtain an evaluation result, and dynamic iterative optimization is carried out on the multi-mode AI diagnosis model group and the structured knowledge graph simultaneously according to the evaluation result to obtain an improved intelligent diagnosis closed loop. Further, according to the heating operation and maintenance data, a structured knowledge graph comprising a physical topology layer, an operation feature layer, a fault mode layer and a regulation strategy layer is constructed, and the method comprises the following steps: The physical topology layer is co