A lithium-ion battery state-of-health prediction model based on physical information constraints and multimodal feature fusion

Abstract

Accurate estimation of lithium battery state-of-health (SOH) is essential for ensuring safe operation and efficient utilization. To address the challenges of complex degradation factors and unreliable feature extraction, we develop a novel SOH prediction model integrating physical information constraints and multimodal feature fusion. Our approach employs a multi-channel encoder to process heterogeneous data modalities, including health indicators, raw charge/discharge sequences, and incremental capacity data, and uses multi-channel encoders to achieve structured input. A physics-informed loss function, derived from an empirical capacity decay equation, is incorporated to enforce interpretability, while a cross-layer attention mechanism dynamically weights features to handle missing modalities and random noise. Experimental validation on multiple battery types demonstrates that our model reduces mean absolute error (MAE) by at least 51.09% compared to unimodal baselines, maintains robustness under adverse conditions such as partial data loss, and achieves an average MAE of 0.0201 in real-world battery pack applications. This model significantly enhances the accuracy and universality of prediction, enabling accurate prediction of battery SOH under actual engineering conditions.

1. Introduction

Lithium-ion batteries, as electrochemical energy storage devices, experience aging processes that are influenced by the coupled effects of multiple factors [1, 2], including temperature, load conditions, and depth of discharge [3, 4]. Moreover, the dynamic nature of their operating environments results in complex degradation pathways and variations in capacity decay rates [5, 6]. The long cycle life and extended duration of aging experiments lead to a scarcity of data for analyzing battery degradation, which complicates the development of precise and generalizable remaining applicable life prediction models and increases the difficulty of accurately predicting battery state-of-health (SOH, Sh). Therefore, extracting practical SOH features and establishing high-precision SOH prediction models mitigate resource waste, extend battery service life, and promote the sustainable development of the industry supply chain [7].

Sensors cannot directly measure SOH, a fundamental indicator in battery management systems (BMS). Thus, accurately estimating and predicting battery SOH using measurable parameters has become a multidisciplinary research hotspot [8, 9]. Batteries’ SOH is typically defined by the capacity decay equation, which is the ratio of the maximum capacity currently deliverable by the battery to its rated capacity [10]. Sh = Caged / Crated × 100% (1) where Caged denotes the current battery capacity, and Crated denotes the rated battery capacity.

Current research on battery SOH prediction can be broadly categorized into experimental methods [11, 12], traditional model-based methods [13 −15], and data-driven methods [16 −19]. Experimental approaches design specific test conditions to obtain characteristic parameters directly reflecting battery aging, thereby analyzing capacity degradation. For instance, GALEOTTI et al [20] used electrochemical impedance spectroscopy to analyze internal resistance and predict SOH. However, these methods cannot be applied to in-service battery packs. Model-based methods simulate battery aging by constructing spatial state equations to forecast degradation paths and estimate SOH. Such models are generally classified into empirical models [21], equivalent circuit models [22, 23], and electrochemical models [24−26]. For example, SINGH et al [27] developed a semi-empirical model for large-capacity ternary lithium batteries to fit voltage curves and determine maximum capacity after each cycle, enabling rapid SOH estimation; ZHANG et al [28] proposed a universal open circuit voltage (OCV)-SOC state equation and parameter identification method insensitive to battery aging. Moreover, another research focus is combining model-based spatial state equations with Kalman filtering algorithms. LIU et al [29] introduced an autoregressive equivalent circuit model and employed a square-root unscented Kalman filter for jointed SOC and SOH estimation, while HU et al [30] established a ...