Dynamic sensor selections for remote prognostics

Abstract

Precise prognostics of operational units is crucial to avoid catastrophic failures. Units in sectors like automotive and aviation, often operating remotely, are equipped with advanced computing systems and wireless sensors that collect operational data. While local computing systems can now autonomously evaluate maintenance needs using server-deployed algorithms, Original Equipment Manufacturers (OEMs) still desire  continuous collection of  the original sensor data to a central server for subsequent data analysis. However, high transmission costs and bandwidth limitations in practice often necessitate selective transmission of sensor data in real time. In addition, dynamic sensor selection must adapt to changing operating environments and accurately characterize the underlying failure evolution. In this study, we leverage the computational capability of local edge units to dynamically select the most informative sensors at the edge unit for transmission in order to achieve accurate prognostic performance. Specifically, we formulate the sensor selection problem as an optimization task that minimizes prediction accuracy loss due to sensor reduction while adhering to the bandwidth constraints. A key challenge in this optimization is that the predictive model is a complex neural network which is challenging to optimize. To address this issue, we propose to approximate the optimization problem within a small neighborhood using linear systems. In this way, the optimization problem can be solved using an efficient dynamic programming algorithm. Validation studies on an aircraft gas turbine engine dataset demonstrate the superiority of our method over other benchmarks.