Exhaust manifold temperature (EMT) plays a critical role in combustion control,
engine performance, and thermal safeguarding of key hardware components in
combustion engines. EMT is a primary power limiting constraint, as exceeding
allowable temperature limits can lead to severe thermal damage to the exhaust
manifold and turbine housing. Accurate estimation of EMT is therefore essential
to ensure engine operation within allowable operating limits. EMT is typically
measured using physical temperature sensors, which are prone to soot fouling,
durability degradation, and calibration drift over time, while also increasing
system cost and complexity. These limitations motivate the development of
reliable virtual sensing approaches capable of replacing or augmenting physical
EMT measurements.
This work presents a machine learning–based virtual sensor for real-time
estimation of exhaust manifold temperature, with a strong focus on production
feasibility within resource constrained engine electronic control units
(ECUs).
Model development, training, and initial evaluation are conducted using data
generated from a high-fidelity one-dimensional engine simulation environment
that captures a wide range of steady state and transient operating conditions.
The trained models are subsequently validated using engine measurement data to
assess robustness and generalization under real world operating scenarios.
Multiple lightweight supervised regression architectures are investigated to
identify the best model that provides an effective balance between estimation
accuracy and embedded implementation efficiency.
Hardware implementation confirmed production feasibility, with the proposed EMT
virtual sensor achieving deterministic real-time execution while requiring only
20% additional CPU utilization and 3.93ms worst-case execution time.
The proposed framework establishes a feasible and scalable pathway for deploying
machine learning–based EMT virtual sensors in engine controllers, enabling
reduced reliance on physical exhaust manifold temperature sensors and supporting
the transition toward intelligent, software defined combustion engine control
systems.