Edge Building Computing for Low-Latency ML Workload Partitioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing building management systems face challenges in efficiently managing computing workloads, optimizing machine learning model performance, and enhancing sustainability, particularly in edge devices, without effective methods for workload distribution, device selection, and performance assessment.
Innovation Solution
A method for processing compute activities in building management systems involves determining the most suitable on-premises or off-premises devices to handle workloads based on characteristics such as computing resources, latency, and sustainability impact, with dynamic edge computing architectures for workload partitioning and machine learning model retraining based on performance assessments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If computing workloads are processed locally on edge building devices, then latency is reduced and responsiveness is improved, but device complexity and resource requirements increase
Solution Approach 1:
The system segments computing workloads into different categories (real-time, near-real-time, batch processing) and distributes them to appropriate devices based on urgency and resource requirements. Critical time-sensitive tasks are processed locally on edge devices, while less urgent tasks are offloaded to cloud or other building devices.
Solution Approach 2:
Building devices are designed to serve multiple functions - they act as both building control devices and computing resources. The same HVAC controller, lighting system, or security device can process local workloads and simultaneously be scheduled as a computing resource for other tasks, eliminating dedicated hardware overhead.
2Productivity
If computing workloads are distributed across multiple building devices, then resource utilization is improved and sustainability is enhanced, but system complexity increases
Solution Approach 1:
The workload assignment system is fully dynamic, continuously monitoring device availability, workload characteristics, and building conditions to optimize task distribution in real-time. Devices can be dynamically added or removed from the computing pool, and workload assignments are adjusted based on changing conditions without manual reconfiguration.
Solution Approach 2:
The system automatically performs workload assignment, performance assessment, and optimization without requiring manual intervention. The workload assignment module autonomously determines optimal device selection and task distribution, while performance monitoring triggers automatic reassignment when devices become unavailable or underperforming.
3Measurement precision
If machine learning models are retrained based on performance assessments, then model accuracy is improved, but computing resources and time are consumed
Solution Approach 1:
The system implements continuous feedback loops where model performance is monitored against actual building conditions and outcomes. When performance degradation is detected through assessment metrics, the system automatically triggers retraining using available device data, creating a closed-loop system that maintains accuracy while minimizing unnecessary retraining computations.
Solution Approach 2:
Rather than continuously retraining models at full computational intensity, the system applies partial retraining only when performance thresholds are breached. The retraining process uses available data from participating devices proportionally, balancing accuracy improvement with resource conservation by avoiding excessive computation when models are still performing adequately.
Data Source
AI summary
A method for enhancing performance of a machine learning model executing on an edge building device of a building includes executing, by one or more processors of the edge building device, the machine learning model, generating, by the one or more processors of the edge building device, an assessment of the performance of the machine learning model on the edge building device, and responsive to the assessment indicating the performance of the machine learning model is below a first level: retraining, by the one or more processors of the edge building device, the machine learning model at the edge building device, and causing, by the one or more processors of the edge building device, a device other than the edge building device to retrain the machine learning model.