On the Limits of End-to-End Learning: Why Statistical Features Dominate in Hyper-Fragmented Battery Prognostics

Fanzhe Fu, Juren Li, Songkang Wen, Jianfeng Zhang, Yang Yang

Keywords: Battery State-of-Health (SOH), End-to-End Learning, Feature Engineering, Hyper-Fragmentation, Real-World Data, Representation Learning, Prognostics and Health Management (PHM)

TL;DR: On noisy, "hyper-fragmented" real-world battery data, a simple model using four engineered statistical features (R^2≈0.80) outperforms complex end-to-end deep learning models (R^2≈0.12), revealing a boundary of the pure end-to-end paradigm.

Abstract: The estimation of State-of-Health (SOH) for Electric Vehicle (EV) batteries from real-world operational data is a critical industrial challenge, primarily due to the "hyper-fragmented" nature of the data. While recent studies have shown that complex hybrid deep learning models, which rely on extensive hand-crafted features, can achieve high performance on this data, a fundamental question remains unanswered: Can the prevailing end-to-end learning paradigm autonomously learn effective representations from such noisy, fragmented raw signals without the aid of domain-specific feature engineering? This paper directly investigates this question through a rigorous comparative study. We contrast two starkly different paradigms: (1) a traditional machine learning approach using a CatBoost model on a novel, compact 4-dimensional statistical feature vector derived from lifetime vehicle signals, and (2) a pure end-to-end approach employing a comprehensive suite of seven advanced deep learning architectures, including 1D-CNNs, LSTMs, and Transformers. Our results reveal a significant performance disparity: the feature engineering approach achieves a robust $R^2$ of approximately 0.80, whereas the best-performing, statistically validated end-to-end model only reaches an estimated $R^2$ of 0.12. This work provides compelling empirical evidence that for high-noise, hyper-fragmented industrial time-series tasks, the information encoded in simple statistical features provides a more effective signal for prognostics than representations learned by current end-to-end architectures, highlighting a critical boundary for their application.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 5057