A Novel Real-Time Framework for Embedded Systems Health Monitoring

Authors:

Abstract
This paper proposes a real-time framework, used as a multi-step ahead forecasting system that can be used for prognostics and health management (PHM), aiming to address a number of industrial-scaled embedded systems where failures can lead to catastrophic events. A novel framework is presented, which combines strong offline training using machine learning techniques with online multi-step forecasting for realtime condition monitoring and prognostics. The offline training is performed with the objective to derive a model capable of predicting the system of interest within fixed performance criteria, which are the metrics that provided the best accuracy and the estimated prediction time to anticipate the performance in real-time. The online inference, on the other hand, uses techniques of multi-step forecasting to provide an advisory window for the operator in case one or multiple faults lead the system for a non-operational state or a hazardous situation. This paper propositions are tested and validated on a real system dataset, where frequent faults were detected during a certain period of time. Dimensionality reduction techniques were initially used on the dataset, ensuring that the overall model accuracy was not impaired, then an optimized bidirectional long short-term memory (LSTM) model was chosen due to its inherent capacity to solve the issue of vanishing gradients for time-dependent and sequential data series. The optimized model is then flashed on two target microcontrollers, which continuously refresh data received from the sensors and actuators in the field. The proposed framework was compared with other strategies, tested and validated on a real application case.

This publication has 32 references indexed in Scilit: