Lithium-ion batteries are the predominant energy storage solutions in various power systems and portable devices owing to their superior energy density, long cycle life, low self-discharge rate, and environmental sustainability. However, aging and performance degradation of lithium-ion batteries during actual use significantly limit their further development. State of charge (SOC) serves as a critical indicator for assessing the remaining capacity and overall health of a battery, making its accurate estimation essential for optimizing battery management systems and enhancing performance. Conventional SOC estimation methods rely primarily on electrical parameters such as voltage and current, which are susceptible to electromagnetic interference and measurement noise, thereby reducing the accuracy of SOC estimation. Therefore, exploring SOC estimation methods based on nonelectrical parameters, particularly optical sensing technologies, has significant practical implications. Fiber Bragg grating (FBG) sensors are a focal point of research in lithium-battery state monitoring owing to their superior measurement accuracy, high resistance to electromagnetic interference, and compact size. The integration of optical sensing technology with deep learning offers a promising avenue for more accurate SOC estimation, presenting significant economic and practical value.

A novel method based on FBG sensors and deep learning models has been proposed to estimate the SOC of lithium-ion batteries. First, a lithium-battery state-monitoring system was established, incorporating FBG sensors and a tunable fiber Fabry?Pérot (FFP) filter. Two FBG sensors were affixed to the surface of the lithium battery to measure changes in the characteristic grating wavelengths resulting from strain and temperature variations during the charge and discharge experiments. Subsequently, a convolutional neural network (CNN) was employed to extract spatial features from the observed wavelength variations. Finally, a gated recurrent unit (GRU) was utilized to establish a deep-learning framework for SOC estimation based on the extracted data features. The accuracy and economic viability of the proposed method were validated by comparing the experimental results with those of conventional methods, demonstrating its advantages in enhancing the SOC estimation accuracy and reducing system costs.

To detail the variations in parameters during the charging and discharging of the battery, we record the voltage, current, and wavelength changes caused by surface strain and temperature, using a battery testing system (BTS) to manage the processes. The current and voltage steps are set, dividing the entire cycle into four distinct stages: constant current discharge, resting, and constant current/constant voltage charging. The changes in the SOC during the charging and discharging of the lithium-ion battery are calculated based on the SOC definition (Fig. 5). Throughout the experiments, the two FBG sensors monitor the changes in the grating wavelength caused by strain and temperature variations. The absolute wavelength drift of the FBG sensors caused by surface strain, and the spectral position of the reference FBG, representing the temperature information, are calculated (Fig. 6). After obtaining the nonelectrical parameters from the FBG, further SOC estimation for the lithium battery is conducted. Using the absolute wavelength drift of the sensing FBG and the spectral position of the reference FBG as input features, a deep learning framework, CNN-GRU, is developed. Accurate SOC estimation for the lithium-ion battery is performed during the charging and discharging cycles, and the evaluation metrics for the estimated results are calculated (Table 1). The experimental results indicate that the prediction model constructed using only nonelectrical input features can accurately predict the SOC of the lithium battery (Fig. 7). Additionally, using the input feature data provided by the FBG sensors, the hybrid model captures the nonlinear relationship between the nonelectrical parameters and the SOC of the lithium battery more effectively, significantly improving SOC estimation.

This study employs two FBG sensors attached to the surface of A123 soft-pack lithium-ion batteries to measure the changes in strain and temperature in real time during charging and discharging. A nonlinear relationship model between optical sensing signals and electrochemical signals is established using deep learning algorithms. Consequently, SOC estimation for the lithium battery is conducted. The results indicate that using only two nonelectrical parameters provided by the FBG sensors yields a relatively accurate estimation of SOC, highlighting the significant potential of nonelectrical parameter features in SOC prediction. Both single and hybrid deep learning models are incorporated into SOC estimation based on FBG sensing. The experimental results show that the hybrid model, CNN-GRU, achieves higher accuracy than the single-model GRU. This provides a more reliable solution for SOC estimation in lithium batteries. The findings address high costs associated with incorporating FBG sensing into lithium-ion battery state estimations. Concurrently, the feasibility of estimating the SOC using pure wavelength signals is validated. This provides valuable insights for extending battery lifespan and improving energy utilization efficiency.