The widespread electrification of transportation and the integration of renewable energy sources have created an unprecedented demand for efficient and reliable energy storage. Among the various technologies, the li ion battery has emerged as the dominant solution due to its superior energy density, relatively long cycle life, and decreasing cost. Its applications span from powering portable electronics and electric vehicles (EVs) to stabilizing electrical grids. The reliable and safe operation of these systems hinges on the precise management of the li ion battery. A core function of any Battery Management System (BMS) is to accurately monitor key states, among which the State of Health (SOH) is paramount for predicting longevity, ensuring safety, and mitigating user “range anxiety.” Direct measurement of a li ion battery’s SOH is impractical. Therefore, indirect estimation using operational data such as voltage, current, and temperature has become the standard approach. In recent years, data-driven methods have gained immense popularity as they bypass the need for modeling complex internal electrochemical and aging mechanisms, instead learning the relationship between measurable data and battery degradation directly. This paper provides a comprehensive review of data-driven SOH estimation methods for li ion batteries, analyzing their principles, comparing their advantages and limitations, and discussing future research directions.
Definition and Influencing Factors of SOH
The SOH of a li ion battery quantifies its degradation level relative to its fresh condition. There is no single universal definition, but it is commonly characterized by capacity fade or internal resistance growth. The end-of-life (EOL) for a li ion battery in automotive applications is typically defined as a 20-30% capacity loss or a 100-160% increase in resistance from its initial value.
Capacity-based SOH: This is prevalent in applications like EVs where range is critical. It is defined as the ratio of the current maximum available capacity to the nominal or initial capacity.
$$SOH_C = \frac{C_{current}}{C_{nominal}} \times 100\%$$
Resistance-based SOH: This is more relevant for high-power applications like hybrid EVs, where power capability is crucial. It is defined based on the growth of internal resistance, often the ohmic resistance.
$$SOH_R = \frac{R_{EOL} – R_{current}}{R_{EOL} – R_{new}} \times 100\%$$
where \(R_{EOL}\) is the resistance at end-of-life, \(R_{current}\) is the present resistance, and \(R_{new}\) is the initial resistance.
The degradation of a li ion battery, leading to SOH decline, is influenced by a complex interplay of factors:
- External Stressors: These include operational conditions such as temperature (high temperatures accelerate side reactions, low temperatures hinder kinetics), charge/discharge rate (high C-rates cause stress), depth of discharge (DOD), and storage voltage.
- Internal Mechanisms: These are the fundamental physical and chemical processes that cause aging, such as the growth and evolution of the Solid Electrolyte Interphase (SEI) on the anode, lithium plating, structural disordering of electrode materials, loss of active material, and electrolyte decomposition.
The data-driven approaches reviewed herein aim to correlate the manifestations of these aging processes (observable in voltage, current, and temperature profiles) with the quantitative SOH metric, without explicitly modeling the underlying mechanisms.
Overview of Data-Driven SOH Estimation Frameworks
Data-driven methods leverage statistical learning and pattern recognition algorithms to establish a mapping between battery operational data (features) and its SOH. The general framework involves two main phases: offline training and online estimation. During offline training, historical aging data from multiple li ion battery cells or packs is used. Health indicators (HIs) or features are extracted from measurable signals like voltage and temperature curves during specific operational segments (e.g., constant-current charge). These features are then fed into a learning algorithm along with the ground-truth SOH (e.g., from capacity tests) to train a model. For online estimation, the same features are extracted from the in-use li ion battery’s data in real-time and fed into the pre-trained model to output an estimated SOH value.
Machine Learning-Based Methods
This category encompasses algorithms that learn complex, non-linear relationships between input features and the target SOH. They are particularly powerful when large datasets are available.
1. Artificial Neural Networks (ANN)
ANNs, inspired by biological neurons, are universal function approximators. They consist of interconnected layers (input, hidden, output) where each connection has a weight. Common variants used for li ion battery SOH estimation include:
- Back-Propagation Neural Network (BPNN): A classic multi-layer perceptron trained using gradient descent. It can model non-linear degradation but risks overfitting and requires careful tuning.
- Feedforward Neural Network (FFNN): Information flows in one direction, from input to output. They are effective but may struggle with temporal dependencies in data.
The strength of ANNs lies in their powerful learning capability and adaptability. However, they often require substantial amounts of training data, and their “black-box” nature provides little insight into the degradation physics of the li ion battery.
2. Support Vector Machines (SVM) and Relevance Vector Machines (RVM)
These are kernel-based methods effective for regression (SVR, RVR) and classification.
- SVM/SVR: Seeks to find a hyperplane that best fits the data in a high-dimensional space defined by a kernel function (e.g., Radial Basis Function). It often yields a sparse model defined only by “support vectors.” It performs well with small to medium-sized datasets related to li ion battery aging.
$$\min_{\omega, b} \frac{1}{2} \|\omega\|^2 + C \sum_{i=1}^{l} (\xi_i + \xi_i^*)$$
$$ \text{s.t. } y_i – (\omega^T \phi(x_i) + b) \le \epsilon + \xi_i^*$$
$$ \quad \quad (\omega^T \phi(x_i) + b) – y_i \le \epsilon + \xi_i$$
$$ \xi_i, \xi_i^* \ge 0, i=1, …, l$$
where \(C\) is a regularization parameter, \(\xi_i\) and \(\xi_i^*\) are slack variables, and \(\epsilon\) defines an insensitive zone. - RVM/RVR: A Bayesian counterpart to SVM, it typically produces an even sparser model. While computationally efficient, its predictions can be sensitive to hyperparameter settings and data fluctuations for a li ion battery.
3. Gaussian Process Regression (GPR)
GPR is a non-parametric, probabilistic model. It not only provides a point estimate for the SOH of a li ion battery but also a measure of uncertainty (prediction variance), which is highly valuable for BMS decision-making. It defines a distribution over functions and infers this distribution from the training data.
Filter-Based Methods
Filtering algorithms are sequential Bayesian estimation techniques ideal for real-time, online SOH tracking. They often work in tandem with a state-space model, where SOH or a related parameter (like internal resistance) is part of the state vector.
1. Kalman Filter (KF) and Its Variants
These are optimal estimators for linear or linearized systems with Gaussian noise.
- Extended Kalman Filter (EKF): The most common variant for non-linear li ion battery models. It linearizes the system dynamics around the current state estimate. Widely used for joint SOC/SOH estimation.
- Unscented Kalman Filter (UKF): Uses a deterministic sampling technique (Unscented Transform) to approximate the state distribution, often providing better accuracy than EKF for strong non-linearities without the need to compute Jacobians.
- Dual/Dual Extended Kalman Filter (DKF/DEKF): Employs two interacting filters: one for state estimation (e.g., SOC) and another for parameter identification (e.g., internal resistance linked to SOH). This is highly effective for a li ion battery as it allows simultaneous tracking of both dynamic and slowly varying parameters.
The standard KF recursion involves a predict and update step:
$$ \hat{x}_{k|k-1} = F_k \hat{x}_{k-1|k-1} + B_k u_k $$
$$ P_{k|k-1} = F_k P_{k-1|k-1} F_k^T + Q_k $$
$$ K_k = P_{k|k-1} H_k^T (H_k P_{k|k-1} H_k^T + R_k)^{-1} $$
$$ \hat{x}_{k|k} = \hat{x}_{k|k-1} + K_k (y_k – H_k \hat{x}_{k|k-1}) $$
$$ P_{k|k} = (I – K_k H_k) P_{k|k-1} $$
where \(x\) is the state, \(P\) is the error covariance, \(K\) is the Kalman gain, \(y\) is the measurement, and \(Q\) and \(R\) are process and measurement noise covariances.
2. Particle Filter (PF)
PF is a Monte Carlo sequential estimation method suitable for highly non-linear and non-Gaussian systems. It represents the posterior probability distribution of the li ion battery’s state (including SOH) using a set of random samples (particles) with associated weights.
$$ p(x_k | y_{1:k}) \approx \sum_{i=1}^{N} w_k^i \delta(x_k – x_k^i) $$
It is particularly powerful for long-term SOH trajectory prediction and Remaining Useful Life (RUL) prognostics, as it can naturally handle uncertainty propagation. However, it is computationally intensive, and its performance depends on the design of the proposal distribution and the risk of particle degeneracy.
Time Series and Deep Learning Methods
These methods explicitly model the temporal dependencies in the li ion battery’s aging data, treating SOH evolution as a sequential process.
1. Autoregressive Integrated Moving Average (ARIMA) Models
ARIMA is a classic linear stochastic model for time series forecasting. It combines Autoregressive (AR), differencing (I), and Moving Average (MA) components. While simple and interpretable, its linear assumption often limits its accuracy for the non-linear capacity fade of a li ion battery, especially for long-term predictions.
2. Recurrent Neural Networks (RNN)
RNNs are a class of neural networks designed for sequential data. They have internal loops (recurrent connections) that allow information to persist, making them suitable for learning temporal patterns in li ion battery degradation data.
3. Long Short-Term Memory (LSTM)
LSTM is a specialized RNN architecture designed to overcome the vanishing gradient problem, enabling it to learn long-term dependencies—a critical aspect for SOH estimation. It uses gating mechanisms (input, forget, and output gates) to regulate information flow. The core equations for an LSTM cell are:
$$ f_t = \sigma(W_f \cdot [h_{t-1}, x_t] + b_f) $$
$$ i_t = \sigma(W_i \cdot [h_{t-1}, x_t] + b_i) $$
$$ \tilde{C}_t = \tanh(W_C \cdot [h_{t-1}, x_t] + b_C) $$
$$ C_t = f_t \odot C_{t-1} + i_t \odot \tilde{C}_t $$
$$ o_t = \sigma(W_o \cdot [h_{t-1}, x_t] + b_o) $$
$$ h_t = o_t \odot \tanh(C_t) $$
where \(f\), \(i\), and \(o\) are the forget, input, and output gate activations; \(C\) is the cell state; \(h\) is the hidden state; and \(\odot\) denotes element-wise multiplication. LSTMs have demonstrated superior performance in capturing the complex, non-linear degradation trajectory of a li ion battery from sequences of operational data.
| Method Category | Typical Algorithms | Key Advantages | Main Challenges/Limitations | Suitability |
|---|---|---|---|---|
| Machine Learning | ANN, SVR, RVR, GPR | High accuracy with good data; models complex non-linearities; no need for physical model. | Requires large, representative datasets; risk of overfitting; “black-box” interpretation. | Offline calibration, cloud-based BMS. |
| Filter-Based | EKF, UKF, PF | Inherently online, real-time capability; provides uncertainty quantification; good for joint state-parameter estimation. | Often requires an underlying state-space model; PF is computationally heavy; performance depends on noise statistics. | Onboard, real-time SOH tracking. |
| Time Series / Deep Learning | LSTM, GRU | Excels at learning temporal patterns and long-term dependencies; very high predictive accuracy. | High computational cost for training; requires large sequential datasets; complex onboard deployment. | Cloud-based prognostics, advanced BMS with significant compute. |
Critical Discussion and Future Perspectives
Data-driven methods have significantly advanced the state of the art in li ion battery SOH estimation. However, several critical challenges must be addressed to transition these techniques from laboratory validation to robust, widespread deployment in real-world applications like electric vehicles.
1. Interpretability and Physics-Informed Learning: Pure data-driven models are often black boxes. Future methods should strive to integrate physical or electrochemical principles with data-driven approaches. For instance, using known aging laws to constrain the learning process or designing neural network architectures that reflect known degradation modes of a li ion battery can improve model robustness, generalizability, and trustworthiness.
2. Generalization Across Diverse Conditions: Most models are trained and validated on data from specific cycling profiles (e.g., constant current) and temperatures. A model’s performance can degrade under the highly dynamic and variable loads seen in real EV driving. Developing methods that are adaptive or inherently generalizable across different usage patterns, temperatures, and even different li ion battery chemistries is crucial. Leveraging transfer learning and meta-learning are promising directions.
3. Feature Engineering and Robustness: The choice of Health Indicators (HIs)—such as voltage curve increments, constant-voltage charging time, or internal resistance—directly impacts performance. These features must be robust to sensor noise, measurement errors, and varying operational conditions. Research into automatic, end-to-end feature learning from raw data (e.g., using 1D convolutional layers) can help mitigate this dependency on manual feature design.
4. Computational Complexity and Onboard Deployment: While sophisticated models like deep LSTMs offer high accuracy, their computational and memory footprints may be prohibitive for current microcontroller-based BMS units. Future work needs to focus on model compression, quantization, pruning, and the development of lightweight yet accurate architectures suitable for embedded implementation in a li ion battery management system.
5. From Cell-Level to Pack-Level Estimation: An EV battery pack comprises hundreds of cells connected in series and parallel. Cell-to-cell variations due to manufacturing tolerances and uneven aging complicate pack-level SOH estimation. Extending data-driven methods to estimate the SOH of individual cells within a pack and subsequently aggregating this to a pack-level SOH, while managing communication and computational loads, remains a significant engineering challenge.
6. Data Availability and Standardization: The performance of data-driven models is intrinsically linked to the quantity and quality of training data. Creating large-scale, publicly available datasets that encompass diverse aging conditions for various li ion battery types is essential for community progress. Furthermore, standardizing validation protocols for SOH estimation algorithms would enable fair and meaningful comparisons.
In conclusion, data-driven SOH estimation for li ion batteries is a vibrant and critical field of research. While methods based on machine learning, filtering, and time-series analysis have demonstrated considerable success, their evolution must focus on enhancing interpretability, robustness under real-world variability, and computational efficiency for practical BMS integration. The convergence of larger datasets, advanced algorithms, and increased onboard computing power will undoubtedly lead to more intelligent, reliable, and predictive health management systems for li ion batteries, ultimately ensuring the safety, performance, and longevity of the energy storage systems that power our modern world.