Exploring Cloud Assisted Tiny Machine Learning Application Patterns for PHM Scenarios

Abstract

Given the diverse deployments of sensor nodes in prognostics and health management (PHM) applications, the use of small form-factor, low-cost and power-efficient microcontrollers (MCUs) has become a practical option for long-term monitoring and front-end data-processing. Hardware advances have enabled small MCU devices to run light-weight machine learning (especially deep neural networks) thereby enabling inference tasks using tiny machine learning (Tiny ML) models executing closer to the data source sensors. Although TinyML like approaches have previously been proposed for some cases in PHM, existing approaches have mainly targeted PHM applications that use single data sources and case-specific models as opposed utilizing prediction models trained from general machine learning frameworks and requiring fusion of multiple distributed data sources. Unfortunately, pure MCUs lack the capacity to conduct such analytics. This work aims to address these limitations by using TinyML deployed at the edge in cooperation with system-level machine learning executing in the cloud. Specifically, we study applications in which sensor data is collected and used to predict system health status and perform remaining useful life regression. We also show how edge MCU devices and cloud computing can be combined and adapted to satisfy diverse requirements, such as latency, power and communication. We also describe the limitations of the current MCU-based deep learning in data-driven prognostics. To the best of our knowledge, this is the first work to systematically investigate the TinyML-Cloud cooperation for data-driven prognostics. We target this as a vision paper and aim to provide a high-level guideline for future PHM application designs involving smart MCU-based decision making.