Auto-creating Anomaly Detection Models via Knowledge Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Efficiently managing and maintaining distributed physical assets across an enterprise campus is challenging due to the difficulty in scaling anomaly detection and prediction models, which requires domain-specific skills and is inefficient when applied to all asset classes.
Innovation Solution
A method for auto-creating anomaly detection or prediction models based on data from a knowledge graph, utilizing node descriptors and relationship descriptors to select model types and generate features, allowing for automated model creation and deployment across various assets and locations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data scientists manually build anomaly detection models for each asset using domain skills, then model accuracy and reliability are improved, but the complexity of deployment and maintenance increases significantly across multiple assets
Solution Approach 1:
The system enables self-service by automatically generating anomaly detection models without requiring data scientists to manually build each model. The automated model generation system uses knowledge graphs and sensor data to autonomously create, deploy, and maintain models across multiple assets, eliminating the need for expert intervention in each model creation process while maintaining detection reliability
Solution Approach 2:
The patent implements universality by creating a single automated model generation platform that serves multiple asset types and locations. The system uses a unified approach with knowledge graphs that can represent different asset classes (chillers, AHUs, etc.) and automatically generates appropriate models for each, allowing one system to handle diverse anomaly detection needs across the entire enterprise campus
2Adaptability or versatility
If manual model building is used for each asset class, then model customization to domain expertise is improved, but productivity and scalability deteriorate when scaling to all assets
Solution Approach 1:
The system performs preliminary action by pre-defining knowledge graphs that encode domain expertise and relationships between assets, sensors, and environmental factors. These knowledge graphs are built in advance and automatically applied when generating models for new assets, allowing the system to leverage pre-prepared domain knowledge without manual customization for each asset while maintaining high productivity
Solution Approach 2:
The patent uses copying by replicating the automated model generation process across all assets. Once the knowledge graph structure and model generation algorithm are established, the system automatically copies this proven approach to every asset type and location, maintaining consistency and adaptability across diverse assets while achieving scale through systematic replication rather than manual customization
3Manufacturing precision
If skilled data scientists are required to build models, then model quality is improved, but loss of time and resources increases due to dependency on specialized personnel
Solution Approach 1:
The system replaces the mechanical system of manual model building by data scientists with an automated computational system. The automated model generation uses algorithms that process sensor data and knowledge graphs to create models programmatically, substituting human expert manual work with automated processes that maintain precision while dramatically reducing the time and resource investment required for model creation and maintenance
Data Source
AI summary
Data from a knowledge graph associated with an enterprise may be obtained. The knowledge graph may include a plurality of node descriptors that indicate locations, assets, and sensor data feeds of the enterprise, and a plurality of relationship descriptors that indicate relationships amongst the locations, the assets, and the sensor data feeds of the enterprise. An anomaly detection or prediction model associated with a selected one of the plurality of node descriptors may be auto-created based on the data from the knowledge graph. In the auto-creation, one of a plurality of model types for the anomaly detection or prediction model may be selected based on an identified node type of the selected node descriptor.