Lithium-ion battery state-of-health prediction using Inception-V4 network optimized by Modified Hiking Optimization Algorithm

Electrochemical impedance spectroscopy (EIS) represents a non-invasive method of measuring state-of-health (SoH) of lithium-ion batteries, but it is difficult to obtain strong degradation measurements because of complicated responses with frequency. The paper presents a new approach of combining Inception-V4 deep learning architecture with a Modified Hiking Optimization Algorithm (MHOA) to obtain precise SoH estimation using only raw impedance spectra. Unlike traditional approaches that rely on equivalent-circuit modeling or complex multidimensional feature extraction, the proposed end-to-end model directly utilizes the complete electrochemical impedance spectrum. The end-to-end network is trained and tested using an available online EIS dataset. EIS spectra were measured for 14,752 experiments for batteries subjected to controlled laboratory environments (25 °C and 45 °C), 6 states of charge, and cycling from 100% to 80% SoH, for which there were 14 logarithmically spaced frequencies from 0.05 Hz to 1000 Hz. The first validation is on a chronologically separated test set, that is, on later cycles of the same cells to test temporal generalization, in which case the test RMSE is 1.15%. On a completely new (not used for training or validation) cell, a test is on cross-cell generalization which results in an RMSE of 0.148%. Outputs from the dual branch network includes circuit like parameter estimations, for which it has been verified to be close to those in the respective equivalent circuit representations. All presented results are from an offline simulation on time-separated data, and no claim is made towards real-time, embedded BMS performance. Using a dataset of 14,752 EIS measurements collected under different temperature conditions and aging stages, the method achieves an overall test RMSE of about 1.15%. It also shows strong capability to generalize to degradation conditions that were not present in the training data. The dual-branch network is capable of providing accurate SoH predictions and interpretable parameters in the form of equivalent-circuit simultaneously, allowing it to be used in a battery management system. The proposed framework shows stable and reliable performance across the evaluated EIS dataset.