Open Access

ARTICLE

Interpreting Electric Vehicle Powertrain Fault Diagnosis Models Using Multimodal Large Language Model-Based Permutation Feature Importance and Leave-One-Feature-Out Importance Analysis Agents

Jaeseung Lee¹, Jehyeok Rew^2,*

1 School of Electrical Engineering, Korea University, Seoul, Republic of Korea
2 Department of Data Science, Duksung Women’s University, Seoul, Republic of Korea

* Corresponding Author: Jehyeok Rew. Email:

(This article belongs to the Special Issue: Intelligent Dynamics Modeling, Predictive Operations & Maintenance, and Control Optimization for Complex Systems)

Computer Modeling in Engineering & Sciences 2026, 148(2), 21 https://doi.org/10.32604/cmes.2026.084165

Received 17 April 2026; Accepted 28 July 2026; Issue published 28 August 2026

Abstract

Accurate and interpretable fault diagnosis of electric vehicle (EV) powertrains is essential for ensuring operational safety, reliability, and efficient maintenance. Undetected faults in key components such as motors, inverters, and batteries can lead to performance degradation and critical system failures. While machine learning (ML)-based fault diagnosis models have demonstrated strong predictive capability using multivariate sensor data, their black-box nature limits practical trust and adoption in real-world EV applications. In particular, understanding how individual sensor variables contribute to diagnostic decisions remains a major challenge. To address this issue, this study proposes a novel interpretability method for EV powertrain fault diagnosis that integrates permutation feature importance (PFI), leave-one-feature-out (LOFO) importance, and multimodal large language model (MLLM)-based analysis agents. Using a random forest-based fault diagnosis model as the predictive backbone, both global and counterfactual importance signals are extracted to characterize variable relevance from complementary perspectives. Dedicated MLLM-based agents are employed to automatically transform numerical tables and visual explanations from PFI and LOFO analyses into structured and domain-aware textual interpretations. An MLLM-based report generation agent synthesizes these interpretations, enabling consistency analysis, discrepancy identification, and integrated diagnostic reasoning. Experimental results on an EV powertrain fault diagnosis dataset demonstrate that the proposed method not only achieves robust fault classification performance but also produces coherent and accurate explanations aligned with physical characteristics of EV powertrains. The results highlight the potential of combining conventional ML with MLLM-based agentic reasoning to deliver trustworthy fault diagnosis systems for next-generation EVs.

---

Keywords

---