h1

TY  - THES
AU  - Jin, Limei
TI  - AI-based simulation of battery system combined with advanced spectroscopy
PB  - RWTH Aachen University
VL  - Dissertation
CY  - Aachen
M1  - RWTH-2024-11389
SP  - 1 Online-Ressource : Illustrationen
PY  - 2024
N1  - Veröffentlicht auf dem Publikationsserver der RWTH Aachen University 2025
N1  - Dissertation, RWTH Aachen University, 2024
AB  - The surge in electric vehicle adoption necessitates the development of high-performance and reliable battery systems. To optimize battery performance and ensure their safe and efficient operation, the Battery Management System (BMS) plays a crucial role. While BMS is commonly utilized at the pack level to oversee the overall health and performance of a battery pack, the recognition of each individual cell with unique operational characteristics within the pack needs more granular and precise management and monitoring at the cell level. Therefore, battery digital twin, which is a real-time, multi-scale virtual representation of the physical battery cell, comes into view. We aim to develop a simplest possible equivalent circuit model-based digital twin. It is designed to accurately reflect the behavior of the actual battery cell under various load conditions. The virtual battery model serves to simulate these conditions and to obtain a wealth of simulated data accordingly. These large amount of data can then be employed by machine learning algorithms to predict the battery's state, providing optimization feedback to enhance the real battery's operation.Equivalent Circuit Modeling (ECM) of Electrochemical Impedance Spectroscopy (EIS) data is a common technique to describe the current state of batteries and can be used as a virtual battery. To characterize ECM parameters in a data supported way, the Gaussian Process Regression (GPR) based Distribution of Relaxation Times (DRT) technique, which simplify EIS data by deconvolution with a suitable kernel, is used to provide the number of distinguishable features based solely on an individual EIS data set. Here a weighted DRT is employed, where weights are tailored based on the frequency-dependent sensitivity of the data, and GPR is utilized to accurately determine these weights. GPR is capable of estimating an suitable weighting matrix from a single data set, potentially enabling automatized DRT inversion without user intervention. The obtained DRT spectrum is then used for the selection of an equivalent circuit model, its initial parametrization, and setting of constraints. Thereby, the ECM parameters that are fitted to experimental EIS data for a single cycle, vary by the State of Charge (SOC). Eventually, by means of the investigated battery it is discussed that using a combination of DRT and ECM, a more physically relevant description of processes in an electrochemical system can be achieved.Unlike SOC, developing a robust and continuous State of Health (SOH) estimation using EIS to build battery digital twins poses a formidable challenge. To bridge the gap between the experiment with data driven techniques that do not rely on fitting of experimental data using a priori models, ECM parameters over a single cycle are expanded in a high-dimensional Chebyshev space. It facilitates not only a mapping of the SOC dependence with robust boundary conditions, but also an extension towards a more abstract SOH description is possible. Due to the long impedance measurement time, the Quasi Monte Carlo (QMC) method can be employed to generate differently aged battery models with limited experimental impedance data. As data becomes available, the space spanning the possible states of a battery can be gradually refined. The developed framework, therefore, allows for the training of big data models starting with very little experimental information and assuming random fluctuations of the model parameters consistent with available data. Battery data, whether obtained through direct measurement or simulation, is inherently complex and high-dimensional. Therefore, detecting anomalies or deciphering the aging pathways from raw data poses significant challenges. To harness effective battery state management, we propose to extract characteristic information from autoencoder latent spaces. First of all, we undertake a comparative analysis of the latent spaces derived from frequency-based impedance data, and from time-based voltage and current data. This comparison provides confidence in the applicability of methods traditionally used for EIS in laboratory environments to the analysis of raw time-series data. Furthermore, in order to estimate realistic battery usage simulation results by incorporating random noise and variations into the data, which is essential for developing robust battery management systems, we compare the latent space trained from ideal sinusoidal data and that from permuted noisy data. The extracted battery features in the latent space were analyzed using Support Vector Machine (SVM) classifiers with both linear and non-linear kernels to segment batteries into three distinct age groups: fresh, aged and damaged cells. The consistent and observable robust aging motion, as depicted by the distribution of classes in the latent space, remains unaffected by the choice of using time-based or frequency-based data, as well as whether the data is ideal or noisy. It underscores the potential and reliability for the development of pseudo-random complex pulse current excitation that can be used for estimating the battery's state, which can then potentially be utilized to inform the development of optimized cell-based load profiles in the future.While EIS offers valuable insights into a battery’s state, real-world battery operation during driving scenarios involves dynamic state changes, where current and voltage signals are far from ideally sinusoidal. To bridge the gap between EIS and real-world driving cycle analysis, we introduce the concept of a stochastic pulse design compatible with the load profile. This approach starts with frequency-based impedance data as a reference and transitions into stochastic pulse signals that act as a weak perturbation in the time domain, thereby providing optimal contrast regrading the battery's state. The systematic generation of stochastic pulse signals is inspired by fractal curves and is limited by constraints of experimental devices. These signals are designed to mimic the variability and unpredictability of real-world battery usage more closely. The current pulse signals are then simulated through the battery models at different SOC/SOH combinations, yielding the corresponding voltage responses. The analysis of this pairwise current/voltage data is also conducted in the latent space of an Autoencoder, which comprises essential features extracted from the input data. Here, traditional EIS data is computed as a reference standard to compare the effects of different pulse sequences against a well-established benchmark in battery analysis. For a visual and quantitative comparison between two latent spaces, Quantile-Quantile (Q-Q) plots are employed to estimate performance among different pulse sequences when replicating the distribution of the EIS-derived latent space. Additionally, their performance can also be visual by latent space segmentation. By correlating the segmentation of latent space with battery aging indicators, we validate the representativeness of our identified best and worst performed pulse signals against conventional EIS data. Furthermore, at certain stages of a battery's aging process, altering the pulse sequence can lead to enhanced performance in terms of age group separation. This insight underscores the potential for adaptive strategies in battery management, where pulse sequences can be dynamically adjusted based on the battery's stage of aging to optimize performance and possibly extend its lifespan.
LB  - PUB:(DE-HGF)11
DO  - DOI:10.18154/RWTH-2024-11389
UR  - https://publications.rwth-aachen.de/record/998435
ER  -