Industrial Knowledge Graph Edge Learning for Real-Time Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems for training AI methods on knowledge graph data require extensive offline processing and are not suitable for real-time implementation on industrial devices with limited computational resources, leading to latency and inefficiencies in data processing and inference.
Innovation Solution
An industrial device and method that enables on-device training and inference for knowledge graphs using a combination of sensors, ETL components, triple stores, and learning components, with neuromorphic hardware for edge learning, allowing continuous learning and dynamic adaptation to data events without the need for external processing servers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If offline processing is used for training AI methods on knowledge graph data, then computational accuracy is improved, but processing time and latency increase significantly
Solution Approach 1:
The system dynamically switches between offline batch processing mode for comprehensive training and online incremental learning mode for real-time updates. The learning component can operate in different modes depending on whether new data arrives, allowing the system to balance between thorough offline analysis and responsive online adaptation, thereby reducing overall processing time while maintaining accuracy.
Solution Approach 2:
The system performs preliminary offline training to establish initial knowledge graph embeddings and models. This preliminary action prepares the system in advance, so when real-time data arrives, the online learning component can quickly incrementally update the pre-trained models without starting from scratch, significantly reducing the latency for real-time predictions.
2Measurement precision
If extensive offline processing is performed, then model training accuracy is improved, but device resource requirements and complexity increase
Solution Approach 1:
The learning system is segmented into distinct components: an offline learning component for batch processing and model training, and an online learning component for real-time incremental updates. This segmentation allows each component to be optimized independently - the offline component can perform comprehensive training while the online component handles lightweight real-time updates, reducing overall system complexity.
Solution Approach 2:
The knowledge graph serves as an intermediary structure that bridges offline and online processing. Offline processing generates and updates the knowledge graph with comprehensive data, which then serves as the foundation for online learning. This intermediary representation allows complex offline analysis to be decoupled from real-time processing requirements.
3Speed
If real-time learning is implemented on industrial devices, then response time is improved, but energy consumption increases
Solution Approach 1:
The system uses periodic batch updates combined with event-driven incremental learning. Instead of continuously processing all data in real-time, the system performs comprehensive learning periodically in batch mode, and only performs lightweight incremental updates when specific data events occur. This periodic action reduces energy consumption compared to continuous real-time processing while maintaining responsive performance.
Solution Approach 2:
The online learning component focuses computational resources only on local updates of specific knowledge graph portions that are relevant to incoming data events, rather than reprocessing the entire knowledge graph. This local quality approach minimizes energy consumption by performing targeted incremental updates only where needed, while maintaining real-time response capability.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
Industrial device (ED) for building and/or processing a knowledge graph (KG), with at least one sensor (S) and/or at least one data source (DS) configured for providing raw data (RD), with an ETL component (ETLC), configured for converting the raw data (RD) into triple statements (T), using mapping rules (MR), with a triple store (ETS), storing the triple statements (T) as a dynamically changing knowledge graph (KG), with a learning component (LC), configured for processing the triple statements (T) in a learning mode, and for performing an inference (IF) in an inference mode, and with a control component (CC), configured for switching between different modes of operation of the learning component (LC).