Basic Study on Applications of LSTM to State Estimation and Diagnosis in Powertrain Control Systems

Abstract

This paper deals with basic application studies of long short-term memory (LSTM) networks to state estimation and diagnosis problems in automotive powertrain control systems. Since an LSTM network can be transformed into a nonlinear state-space equation, it is expected that nonlinear behavior of the powertrain systems can be expressed with it. The LSTM networks are learned with the data generated by an engine simulator built with GT- Power and applied to a diesel engine air path system to estimate turbocharger rotation speed, exhaust manifold pressure, NOx and soot emissions. Moreover, a diagnosis of characteristic variation of the turbocharger is considered based on an LSTM network. Numerical examples demonstrate the effectiveness of the present method.