Lithium-ion batteries have emerged as the primary power source for electric mobilities. To ensure the safe operation of the batteries throughout their lifespan, accurate remaining useful lifetime (RUL) prediction is required. This paper proposes ensemble strategies that integrate two different Gaussian process regression (GPR) methods and model-based methods to enhance the robustness of the prediction. The first GPR is based on the forward extrapolation of the measured capacity sequence. The second GPR is based on the extrapolation of the measured feature and then inputs the predicted feature into a capacity estimation model. The first ensemble strategy is the weighted ensemble method, which uses the least squares method to determine the weighted coefficients. The second strategy is a more conservative method, which chooses the fastest degradation path between two basic methods at each prediction step. The third strategy is particle filter (PF) to combine the predicted data from different methods. The batteries aged by a real forklift aging profile and open access dataset are used to verify the proposed methods. The results of all methods based on different percent of data are analyzed. The results show that individual methods may obtain different prediction results, while ensemble strategies have accurate and robust predictions. The PF for capacity-based and feature-based methods has the best performance with the absolute error of RUL less than 23 full equivalent cycles, error of prediction steps less than 1, and negligible simulation time for forklift dataset.

Lithium-ion batteries have been widely applied in e-mobilities and energy storage devices. Fast and accurate state of health (SOH) estimation is crucial to ensure the reliable operation and timely maintenance of these devices. This work proposed multiple linear regression (MLR) models to estimate the SOH of battery applied in forklift load profile and compared the estimation accuracy between extracting features from complete discharging-charging voltage curves and only charging voltage curves. Two features were extracted from two kinds of voltage curves respectively firstly, and the third feature was then extracted from many-step voltage curves to improve the generalization performance. The MLR was used to build the relationship between SOH and features. Finally, root mean square error (RMSE) was employed to evaluate the model accuracy. Results show that the MLR can effectively estimate SOH based on the three features and the estimation accuracy is higher when extracting features from only charging voltage curves.

This work estimates SOH of lithium-ion batteries, aged by a forklift driving profile, based on multiple linear regression and compares the estimation accuracy at three levels. Unlike previous research, this work uses dynamic and field data rather than public datasets. The influence of data size and the position to extract features on the SOH estimation accuracy was researched. It is found that extracting features from smaller voltage segments contains more information. The estimation accuracy can be improved by 24.5% MAPE after the Box-Cox transformation.

The EIS can be effectively applied in Proton exchange membrane fuel cell (PEMFC) for online diagnosis of the system by machine learning methods to prolong the lifetime. The development of advanced EIS techniques lays a solid foundation for the flexibility of data collection based on EIS for the construction of diagnosis model in PEMFC. Feature extraction from EIS can be conducted using raw impedances and based on different interpretation methods of EIS. Different machine learning models show great possibilities for PEMFC diagnosis combining with EIS. Classical classification, extreme learning, simple fuzzy logic, and ensemble learning (without deep learning) can be adopted using countable 1D features of EIS which are interpretable, such as geometric information, ECM parameters, and DRT characteristic frequencies and peaks. Complex deep learning methods can adopt raw EIS for automatic feature extraction. Typically, CNNs can be used to convolve 2D EIS data, processing the real and imaginary parts across the frequency range to extract features automatically. Since impedance spectra are collected at regular or specific time points during cell operation for real-time diagnosis, RNN methods like LSTM can also be applied using features extracted from EIS in theory. The critical steps in developing data-driven diagnosis strategies for PEMFC with EIS and machine learning methods come from seven aspects, namely, determining applicable faults for diagnosis, achieving fast and reliable online EIS data collection, extracting potential features for constructing diagnosis models, expanding the choice of machine learning methods for online FDI of PEMFC based on EIS, conducting comprehensive performance evaluations of diagnosis models, making full use of big data for diagnosis, and diagnosing special faults in high-temperature PEMFCs. EIS combined with machine learning methods can also be applied for online prediction of fuel cell health index, such as ECSA and hydrogen crossover, which are not available by onboard sensors. This work hopes to provide inspirations and new insights for future research on using AI methods in fuel cell health management. Figure 1

It is challenging and valuable to estimate the state of health (SOH) and predict the remaining useful lifetime (RUL) of lithium-ion batteries under random, dynamic, and real operation conditions. This work aims to estimate the SOH of a battery aged using a realistic forklift load profile. Furthermore, the influence on the SOH estimation of the feature extraction and selection based on different positions of one recharging curve was investigated and the RUL of the battery was predicted based on the limited capacities. Filtering methods and support vector regression (SVR)-based subset selection methods were used in sequence to select the final features for battery SOH estimation. The impact of features on the output of SVR-based SOH estimation was analyzed through Partial dependence plots (PDP). The results showed that the proposed feature selection can reduce SOH estimation error from values larger than 5.43% to less than 3.65%. Estimation accuracy is below 1%RMSE based on one sub-curve even the data in the later aging stage not involved in training. Gaussian process regression (GPR) combined with random sampling was used to predict the RUL and uncertainty. The minimum absolute error of RUL prediction was 0.379 FECs.