Logo image
Multi-objective feature selection using HPWOA for improved BMS fault diagnosis in electric vehicles
Journal article   Open access   Peer reviewed

Multi-objective feature selection using HPWOA for improved BMS fault diagnosis in electric vehicles

Buasa Andy Mayingi, Bonginkosi A. Thango and Daniel Okojie
World electric vehicle journal, Vol.17(7), p.359
01/07/2026
Handle:
https://hdl.handle.net/10210/520662

Abstract

Engineering, Electrical & Electronic Science & Technology Transportation Science & Technology Engineering Technology Transportation
Battery management systems (BMSs) in electric vehicles (EVs) are instrumented with an increasing number of heterogeneous sensors, many of which contribute redundant or noisy measurements that increase computational cost without improving diagnostic accuracy. This paper proposes a Binary Hybrid Particle Whale Optimization Algorithm (BHPWOA) for multi-objective feature selection targeting three-class BMS fault diagnosis: OK, Warning, and Critical. The method is evaluated using an 18-feature EV charging dataset with n = 500 samples. BHPWOA encodes candidate feature subsets as binary masks in a continuous [0, 1] position space. It executes a Binary Particle Swarm Optimization (BPSO) phase during the first 50 iterations to rapidly identify a promising subset region, then transfers the global-best mask as the Whale Optimization Algorithm (WOA) leader for the remaining 50 iterations of bubble-net exploitation. A multi-objective fitness function simultaneously penalises classifier error and subset size, directly optimising the accuracy-cost trade-off. BHPWOA selects four features out of 18, corresponding to a 77.8% reduction, and achieves accuracy = 0.710 and macro-F-1 = 0.4455 on the held-out test set. It outperforms all-feature KNN (F-1 = 0.2997), standalone BPSO with six selected features (F-1 = 0.4603), BWOA with two selected features (F-1 = 0.4026), and BSFSA with five selected features (F-1 = 0.4216) on the Pareto-dominant combined fitness objective. The selected subset {CellVoltage(V,) ChargeCurrent(A), SOC%, ChargePower(kW)} achieves the best fitness score of-0.5555, enabling a 77.8% sensor-cost reduction while improving fault detection. Stability analysis across five independent random seeds confirms a mean feature count of 4.0 +/- 0.7 and a mean macro-F1 of 0.441 +/- 0.021, demonstrating algorithmic robustness.
pdf
Research - 2026-08-17T111854.6665.03 MBDownloadView
Open Access CC BY V4.0
url
https://doi.org/10.3390/wevj17070359View
Published (Version of record) Open

Metrics

1 Record Views

Details

Logo image