Hierarchical Edge ML Model Aggregation for System Monitoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems in process industries lack efficient, operator-independent performance monitoring and root cause analysis capabilities due to the complexity and resource-intensive nature of machine learning (ML) model generation, deployment, and maintenance, which hinders the adoption of ML for autonomous system operation.
Innovation Solution
A method for efficient performance monitoring in hierarchical networks of distributed devices, where local ML models from client edges are aggregated by a master edge to generate a global ML model for system monitoring, utilizing distributed online learning to reduce the effort of model generation and adaptation, and enabling real-time performance and condition monitoring.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are deployed for autonomous system operation, then performance monitoring and root cause analysis capabilities are improved, but the complexity of infrastructure and expertise requirements increases
Solution Approach 1:
The patent segments the ML model execution across distributed edge devices in the hierarchical network. Each edge device maintains local models that operate independently, while only necessary model parameters are aggregated at higher levels. This segmentation eliminates the need for a centralized complex infrastructure while maintaining autonomous monitoring capabilities at each device level.
Solution Approach 2:
The edge devices perform self-learning and self-monitoring using locally stored sensor data and pre-trained models. The devices autonomously execute performance monitoring and root cause analysis without requiring continuous external intervention or complex centralized management infrastructure, thereby reducing overall system complexity.
2Extent of automation
If machine learning models are generated and maintained, then autonomous operation capability is improved, but the time and resources required for model generation, adaptation, and retraining increase
Solution Approach 1:
The patent employs pre-trained ML models that are deployed at edge devices before autonomous operation begins. These pre-trained models immediately enable autonomous monitoring and analysis capabilities without requiring time-consuming on-site training or adaptation, allowing rapid deployment of autonomous operations.
Solution Approach 2:
Instead of creating entirely new models for each edge device, the patent uses copies of pre-trained models that can be rapidly deployed. The models are replicated across the hierarchical network, and only necessary parameter updates are transmitted, significantly reducing the time and resources required for model deployment and adaptation compared to training new models from scratch.
3Adaptability or versatility
If distributed ML models are aggregated to create global models, then system-wide performance monitoring is improved, but the communication and processing overhead increases
Solution Approach 1:
The patent extracts only the essential model parameters and aggregated statistics from individual edge devices and transmits them to higher levels in the hierarchical network. By taking out only the necessary information rather than transmitting complete models or all raw data, the system achieves system-wide monitoring capability while minimizing communication and processing overhead.
Data Source
AI summary
A method for system monitoring in a hierarchical network of distributed edge devices includes a master edge, first and second client edges connected via a first communication interface to the master edge, the method including receiving sensor data from a sensor device via a second communication interface, determining a first local model parameter representing a machine learning (ML) model of the at least first client edge based on the sensor data; storing the first local model parameter in a data storage of the at least first client edge; collecting the first local model parameter from the at least first client edge; and generating a global ML model based on the at least first local model parameter, wherein the global ML model is used for monitoring a system performance or a condition of the system.