Interpretable machine learning models for predicting and explaining vehicle fuel consumption anomalies

Barbado, Alberto; Corcho, Óscar · 2022 · Crossref

DOI: 10.1016/j.engappai.2022.105222

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of not only detecting anomalous fuel consumption in vehicle fleets but also explaining the causes behind these anomalies to enable actionable optimization. While existing methods can identify outliers, they often fail to provide the contextual insights required by fleet managers and operators to reduce costs. The authors propose a complete methodology that combines unsupervised anomaly detection with interpretable machine learning (IML) models, specifically Generalized Additive Models (GAMs), to generate feature-relevance explanations. The goal is to align these explanations with domain knowledge and user-specific needs, thereby facilitating prescriptive recommendations for fuel reduction. The study utilizes real-world telematics data from diesel and petrol vehicles across various industrial fleets. The process begins with preprocessing raw IoT data into daily aggregated records, filtering out non-representative trips and handling missing values. Anomaly detection is performed using an unsupervised Box-Plot approach, which establishes dynamic thresholds for fuel consumption based on vehicle categories and route types. To explain the detected anomalies, the authors benchmark three interpretable models: Explainable Boosting Machine (EBM), a proposed variation of EBM (EBM var) that better handles categorical features, and Constrained Generalized Additive 2 Model with Consideration of Higher-Order Interactions (CGA2M+). These models are evaluated using both standard performance metrics and specific Explainable AI (XAI) metrics, including representativeness, fidelity, stability, contrastiveness, and consistency with prior beliefs. The results demonstrate that the proposed methodology effectively identifies fuel consumption anomalies and provides clear, interpretable explanations regarding the contributing features, such as driving behavior and vehicle status. The evaluation reveals that CGA2M+ and the modified EBM offer robust performance while maintaining high interpretability. Crucially, the system generates tailored recommendations for two distinct user profiles: fleet operators receive specific advice for individual vehicles, while fleet managers receive broader insights on driving behavior impacts. The study reports that these actionable insights can lead to potential fuel reductions of approximately 35%. The significance of this work lies in its integration of XAI into the anomaly detection lifecycle, moving beyond binary classification to provide understandable and prescriptive insights. By validating the use of CGA2M+ on real-industry data and proposing metrics to quantify explanation quality, the paper contributes to the field of Responsible AI. It demonstrates that combining interpretable models with domain knowledge and user-centric design can effectively bridge the gap between data-driven anomaly detection and practical operational improvements in fleet management.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• telematics crash prediction