MACHINE LEARNING FOR HETEROGENEOUS DIAGNOSTIC FAULT FORECASTING MODEL IN ELECTRIC VEHICLES

Abstract

Predictive maintenance of electric vehicles (EVs) has gained increasing traction in research around the world, but the literature is scattered and not well integrated in methodological approaches, and not widely used in stratified analysis of operationally diverse vehicle fleets. Studies have been conducted that report on the performance of a vehicle fleet which typically comprises diverse vehicles, each having dramatically different usage profiles, and consequently, different performance numbers that mask the impact of usage levels on component wear and Diagnostic Trouble Code (DTC) generation. This systematic review collates evidence from 77 peer reviewed studies published from January 2015 to April 2026, identified using a search of IEEE Xplore, ScienceDirect, ACM Digital Library and Google Scholar, conducted in March/April 2026. The review is based on four objectives: (i) to evaluate the state of machine learning methods applied to EV DTC forecasting; (ii) to assess the extent to which existing studies account for heterogeneous usage intensity profiles in their analysis and performance reporting; (iii) to identify sensor-level telemetry precursor patterns documented in prior work; and (iv) to assess the transferability of current methodological frameworks to data-scarce emerging EV markets such as Kenya. Given the diversity of metrics and experimental designs reported, the narrative synthesis approach is used in the review, which is done according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines. Results show that sequential deep learning architectures (such as Long Short-Term Memory (LSTM) network and Convolutional Neural Network-LSTM hybrid) always beat ensemble tree-based architectures and classical statistical baseline models with macro-averaged F1-scores of 0.76 to 0.92. Only 77 of which studies included an analysis of usage-stratified performance (5.2%) and none conducted a systematic comparison of multi-class categorical DTC forecasting based on four or more distinct usage intensity profiles simultaneously. Only two reports of lead-time distribution characterisation could be found, and Serradilla et al. (2022) state that a lack of data is the biggest structural obstacle for implementing predictive maintenance in new and emerging EV fleet scenarios. The four substantive research gaps identified are lack of usage-stratified analysis, predominant emphasis on binary fault classification analysis over multi-class DTC categorical forecasting, lack of lead-time distribution characterisation, and absence of empirically validated methodological guidance for data-scarce emerging markets. These gaps inspire the subsequent empirical study that will create and evaluate machine learning models for four use intensity profiles for a five-year, 172,800 EV telemetry observation public dataset.