Author ORCID Identifier

https://orcid.org/0009-0002-4854-2872

Semester

Summer

Date of Graduation

2026

Document Type

Thesis

Degree Type

MS

College

Statler College of Engineering and Mineral Resources

Department

Mechanical and Aerospace Engineering

Committee Chair

Andrew Nix

Committee Member

Brian Woerner

Committee Member

Derek Johnson

Abstract

Battery electric vehicles (BEVs) rely on accurate thermal management to ensure component performance, efficiency, safety, and long-term durability. Critical propulsion system components, including electric motors and high-voltage battery packs, are commonly monitored using physical temperature sensors. However, these sensors increase system cost, introduce additional hardware complexity, and may be impractical for directly measuring temperatures at critical internal locations. Virtual temperature sensors (VTSs) provide a software-based alternative by estimating component temperatures using readily available vehicle operating data.

This research presents the development and validation of a long short-term memory (LSTM) neural network-based virtual temperature sensor capable of simultaneously estimating the motor rotor and high-voltage battery temperatures of a dual motor BEV. Experimental data were collected from the West Virginia University EcoCAR Cadillac LYRIQ prototype vehicle using an onboard controller area network (CAN) data acquisition system. The resulting dataset consisted of 21 on-road vehicle operation events spanning approximately 34.9 hours and more than 1.2 million synchronized data samples. Fourteen vehicle-level signals describing vehicle operating conditions and powertrain performance were used as network inputs, while the three component temperatures served as prediction targets. Data preprocessing included signal decoding, trimming, synchronization, normalization, and feature preparation prior to network training and validation.

The proposed LSTM network was evaluated using both open-loop and closed-loop prediction strategies on previously unseen validation data. Open-loop prediction demonstrated superior performance compared to closed loop, achieving a root mean squared error (RMSE) of 0.67 °C and a mean absolute error (MAE) of 0.29 °C across the validation datasets. The model accurately tracked motor rotor temperatures under both transient and steady-state operating conditions, while battery temperature predictions captured overall thermal trends with slightly reduced accuracy. Although the prediction accuracy demonstrates the effectiveness of the proposed virtual sensing approach, inference times remain above the requirements for real-time embedded implementation, indicating that additional model optimization is necessary prior to in-vehicle deployment.

The results demonstrate that deep learning-based virtual temperature sensing is a viable approach for estimating critical thermal states in BEVs using existing vehicle signals. This work establishes a foundation for future development of intelligent software-based sensing technologies that can improve thermal management while reducing dependence on additional physical sensors.

Download

Share

COinS