The accurate estimation of the State of Health (SOH) for lithium-ion batteries is paramount to ensuring the safety, reliability, and longevity of energy storage systems, particularly in electric vehicles. The degradation of a lithium-ion battery is a complex, nonlinear electrochemical process, leading to diminished capacity and increased internal resistance, which directly impacts system performance and safety. Our study presents a data-driven SOH estimation methodology that synergistically combines Incremental Capacity Analysis (ICA) for robust feature extraction with a Genetic Algorithm-optimized Backpropagation (GA-BP) Neural Network for high-fidelity modeling, successfully overcoming common limitations of traditional machine learning approaches.
The operational health of a lithium-ion battery is most fundamentally tracked by its SOH, commonly defined as the ratio of its current maximum available capacity to its nominal capacity. This intrinsic characteristic of the lithium-ion battery fades over cycles due to irreversible side reactions like solid electrolyte interphase (SEI) layer growth, lithium plating, and active material loss. Consequently, developing an accurate, robust, and computationally efficient SOH estimation algorithm is a critical research frontier for battery management systems (BMS). While model-based approaches offer insights, their reliance on precise electrochemical parameters and susceptibility to model mismatch in real-world conditions pose significant challenges. Alternatively, data-driven methods, empowered by machine learning, learn the complex mapping between measurable battery signals and SOH directly from operational data, offering a powerful and flexible solution for lithium-ion battery management.
Our proposed framework consists of two pivotal stages. First, we process raw voltage-capacity data through the ICA transform to amplify subtle aging signatures hidden within the voltage plateaus characteristic of lithium-ion battery chemistry. Second, we employ a GA-BP neural network to learn the nonlinear relationship between these extracted ICA features and the battery’s SOH. The integration of the Genetic Algorithm is crucial, as it optimizes the initial weights and biases of the BP network, mitigating its well-known drawbacks of slow convergence and proneness to local minima, thereby enhancing the predictive accuracy and robustness for lithium-ion battery SOH estimation.
1. Feature Engineering via Incremental Capacity Analysis (ICA)
Directly using voltage or current profiles from a lithium-ion battery for SOH estimation is suboptimal due to the flat voltage plateaus during charge/discharge, where changes are minimal and noisy. Incremental Capacity Analysis serves as a powerful signal processing tool to transform these subtle changes into prominent, analyzable peaks. The ICA curve, representing the derivative of capacity with respect to voltage, is calculated as:
$$ IC = \frac{dQ}{dV} = \frac{I \cdot dt}{dV} = I \cdot \frac{dt}{dV} $$
Where \( Q \) is capacity, \( V \) is terminal voltage, \( I \) is current, and \( t \) is time. In practice, for a discrete dataset from a lithium-ion battery test, this is computed over fixed, small voltage intervals \( \Delta V \). For our study, \( \Delta V \) was set to 0.05V. This derivative operation effectively magnifies the phase transformation points of the electrode materials within the lithium-ion battery, which shift in a deterministic manner as the battery degrades.
Raw IC curves are inherently noisy due to sensor inaccuracies and operational fluctuations. Therefore, a smoothing filter is essential. We applied a Gaussian filter, which performs a weighted average of data points within a defined window. The Gaussian kernel is defined by:
$$ G(x) = \frac{1}{\sigma\sqrt{2\pi}} \exp\left(-\frac{(x – \mu)^2}{2\sigma^2}\right) $$
Where \( \mu \) is the mean (center of the window) and \( \sigma \) is the standard deviation controlling the smoothing intensity. The application of this filter effectively removes high-frequency noise while preserving the essential peak morphology of the lithium-ion battery’s IC curve, as illustrated in the processed data. The evolution of these smoothed IC peaks across the lifespan of a lithium-ion battery provides a clear visual fingerprint of its aging progression.
To translate this visual evolution into quantifiable model inputs, we extract specific health feature factors from the ICA curve. Observing the curve evolution, the voltage region between approximately 3.8V and 4.1V shows highly systematic peak shifts with aging. We select a subset of IC values at fixed voltage points within this sensitive region as our feature vector. To maintain a low-dimensional yet informative input space, we employ an equal-interval interpolation method, extracting the \( dQ/dV \) values at voltages: 3.85V, 3.90V, 3.95V, 4.00V, 4.05V, and 4.10V.
The relevance of each extracted feature to the actual SOH degradation of the lithium-ion battery must be rigorously assessed. We use the Pearson correlation coefficient \( r \) to quantify the linear dependence between each feature time series and the corresponding SOH time series. For two data series \( x \) and \( y \), the coefficient is calculated as:
$$ r_{xy} = \frac{s_{xy}}{\sqrt{s_{xx} \cdot s_{yy}}} = \frac{\sum_{i=1}^{n} (x_i – \bar{x})(y_i – \bar{y})}{\sqrt{\sum_{i=1}^{n} (x_i – \bar{x})^2 \sum_{i=1}^{n} (y_i – \bar{y})^2}} $$
$$ \bar{x} = \frac{1}{n}\sum_{i=1}^{n} x_i, \quad \bar{y} = \frac{1}{n}\sum_{i=1}^{n} y_i $$
Where \( s_{xx} \) and \( s_{yy} \) are sample variances, and \( s_{xy} \) is the sample covariance. A higher absolute value of \( r \) (closer to 1) indicates a stronger linear correlation, suggesting the feature is a good health indicator for the lithium-ion battery. The correlation analysis for cells from a public dataset is summarized below.
| Cell ID | Feature #1 (3.85V) | Feature #2 (3.90V) | Feature #3 (3.95V) | Feature #4 (4.00V) | Feature #5 (4.05V) | Feature #6 (4.10V) |
|---|---|---|---|---|---|---|
| Cell A | 0.97 | 0.99 | 0.99 | 0.99 | 0.97 | 0.94 |
| Cell B | 0.96 | 0.97 | 0.99 | 0.99 | 0.95 | 0.84 |
| Cell C | 0.98 | 0.98 | 0.98 | 0.97 | 0.89 | 0.67 |
| Cell D | 0.82 | 0.80 | 0.88 | 0.85 | 0.67 | 0.48 |
Based on the correlation analysis, Features #1 through #4 consistently show very high correlation coefficients (mostly >0.95) across multiple cells, confirming their strong representation of lithium-ion battery degradation. Features #5 and #6, while still correlated, show more variability and lower values for some cells. Therefore, to ensure model robustness and input efficiency, we select the first four ICA-derived features (at 3.85V, 3.90V, 3.95V, and 4.00V) as the definitive health feature vector for the subsequent SOH estimation model of the lithium-ion battery.
2. Architecture of the Hybrid GA-BP Neural Network Estimator
2.1 Backpropagation (BP) Neural Network Foundation
The Backpropagation Neural Network is a quintessential multilayer feedforward network renowned for approximating complex nonlinear functions. For estimating the SOH of a lithium-ion battery, we design a three-layer network. The basic computational unit, the neuron, performs a weighted sum of its inputs, adds a bias, and passes the result through an activation function \( f(\cdot) \):
$$ \text{net}_j = \sum_{i=1}^{n} \omega_{ij} x_i + b_j $$
$$ y_j = f(\text{net}_j) $$
Where \( \omega_{ij} \) is the weight connecting input \( i \) to neuron \( j \), \( x_i \) is the input, \( b_j \) is the bias (threshold), and \( y_j \) is the output. For our lithium-ion battery SOH estimator, the network structure is defined as follows:
- Input Layer: 4 nodes, corresponding to the four selected ICA health features.
- Hidden Layer: The number of nodes \( P \) is determined empirically. A common heuristic is \( P = \sqrt{m + n} + a \), where \( m \) is the number of input nodes (4), \( n \) is the number of output nodes (1), and \( a \) is a constant between 1 and 10. We set \( a=8 \), resulting in \( P = \sqrt{4+1} + 8 \approx 10 \). The hyperbolic tangent sigmoid (tansig) function is chosen as the hidden layer activation for its strong gradient.
- Output Layer: 1 node, producing the estimated SOH value (range 0-1). A pure linear activation function is used here.
The network is trained by minimizing the Mean Squared Error (MSE) between its predictions and the true SOH values via gradient descent. However, the performance of a standard BP network is highly sensitive to the initial random values of weights and biases, often leading to slow convergence or trapping in poor local minima—a significant issue when modeling the intricate aging dynamics of a lithium-ion battery.
2.2 Genetic Algorithm (GA) for Global Optimization
The Genetic Algorithm is a metaheuristic optimization algorithm inspired by natural selection. We employ GA to find a superior set of initial weights and biases for the BP network before the gradient-based training commences, thus guiding the search towards a more favorable region in the parameter space. The GA optimization process involves the following steps, meticulously applied to our lithium-ion battery SOH estimation problem:
1. Chromosome Encoding & Population Initialization: All parameters of the BP network (weights \( W_1, W_2 \) and biases \( B_1, B_2 \)) are concatenated into a single real-valued vector, forming one chromosome. A population of \( N \) such chromosomes is randomly initialized. We set the population size \( N = 20 \).
2. Fitness Function Evaluation: The fitness of each chromosome dictates its probability of being selected for reproduction. For our task, lower fitness is better. We define the fitness \( F \) as the sum of absolute errors (SAE) when the network, configured with the chromosome’s parameters, processes the training data:
$$ F = k \cdot \sum_{i=1}^{M} | \hat{y}_i – y_i | $$
where \( \hat{y}_i \) is the network output (estimated SOH), \( y_i \) is the true SOH, \( M \) is the number of training samples, and \( k \) is a scaling coefficient.
3. Selection: We use the roulette wheel selection method. The selection probability \( p_i \) for an individual \( i \) is inversely proportional to its fitness relative to the total population fitness:
$$ p_i = \frac{F_i^{-1}}{\sum_{j=1}^{N} F_j^{-1}} $$
This ensures chromosomes yielding lower SOH estimation error have a higher chance of being selected to pass their “genetic material” to the next generation.
4. Crossover: Selected parent chromosomes exchange genetic information to produce offspring. We use arithmetic crossover. For two parent chromosomes \( a \) and \( b \) at gene position \( j \), the offspring genes \( a’_j \) and \( b’_j \) are:
$$ a’_j = \beta a_j + (1-\beta) b_j $$
$$ b’_j = \beta b_j + (1-\beta) a_j $$
where \( \beta \) is a random number in [0, 1]. The crossover probability is set to \( P_c = 0.7 \).
5. Mutation: Mutation introduces random changes to maintain population diversity and explore new regions of the solution space. We employ a non-uniform mutation operator. The mutation probability for a gene is adaptively tuned based on its fitness:
$$ p_m = \begin{cases}
\alpha \cdot (0.5 – f_d), & 0 \leq f_d \leq 0.5 \\
\alpha \cdot (1 – f_d)^2, & 0.5 \leq f_d \leq 1
\end{cases} $$
where \( f_d \) is the normalized fitness of the individual and \( \alpha = 0.1 \) is a mutation rate coefficient. This adaptive approach protects high-fitness solutions while allowing exploration.
The GA iterates through selection, crossover, and mutation for a predefined number of generations (we set it to 50). The chromosome with the best (lowest) fitness after the final generation is decoded to provide the optimized initial weights and biases for the BP network dedicated to lithium-ion battery SOH prediction.
2.3 Integrated GA-BP SOH Estimation Model Workflow
The complete algorithm for lithium-ion battery SOH estimation synergizes the feature extraction and model optimization stages. The workflow is executed as follows:
- Data Preparation: Split the entire aging dataset of the lithium-ion battery (e.g., cycles 1 to N) into a training set (first 50% of cycles) and a testing set (remaining 50%).
- Feature Extraction: For each cycle in both sets, compute the smoothed ICA curve and extract the four predefined health feature values.
- GA Optimization:
- Define the BP network structure (4-10-1).
- Run the GA for 50 generations to evolve the optimal initial network parameters (weights/biases).
- The fitness function is evaluated on the *training set* only.
- BP Network Training: Initialize the BP network with the GA-optimized parameters. Train the network on the training set using standard backpropagation (learning rate = 0.1, max epochs = 100) until the performance on a validation subset converges.
- SOH Estimation & Validation: Feed the health features from the unseen *testing set* into the fully trained GA-BP model. The output is the estimated SOH for the lithium-ion battery. Compare these estimates with the true values to evaluate model performance.
The synergy is clear: the ICA provides physically meaningful, low-dimensional inputs that characterize the internal state changes of the lithium-ion battery, while the GA-BP network provides a powerful, globally optimized nonlinear regressor to map these inputs to a precise SOH value, overcoming the pitfalls of conventional single-method approaches.
3. Experimental Validation and Comprehensive Results Analysis
The proposed methodology was validated using a publicly available lithium-ion battery aging dataset. Cells were cycled under controlled laboratory conditions with constant-current constant-voltage (CCCV) charging and constant-current (CC) discharging until their capacity faded to 70% of the nominal value. The data includes voltage, current, and capacity measurements for every cycle, providing the ground truth for SOH calculation.
3.1 GA-BP Model Training Performance
The training process demonstrates the efficacy of the hybrid approach. The GA optimization phase efficiently searches the parameter space. The fitness value of the best chromosome in the population decreases rapidly over generations, typically converging to a low-error plateau within 20-30 generations. This indicates the GA successfully finds a high-quality starting point for the lithium-ion battery SOH estimation model.
Subsequently, the BP network training initialized with these optimized parameters exhibits excellent convergence properties. The Mean Squared Error (MSE) on the training, validation, and test subsets decreases sharply in the initial epochs and stabilizes at a very low value (often below \( 1 \times 10^{-4} \)) within about 20 epochs. The close alignment of error reduction across all subsets confirms that the model is learning generalizable patterns related to lithium-ion battery degradation rather than merely memorizing the training data, thus mitigating overfitting.
3.2 SOH Estimation Results on Test Data
The final trained GA-BP models were applied to the independent test data (the latter 50% of cycle life for each lithium-ion battery). The estimation results are highly accurate. The predicted SOH curves closely track the experimentally measured SOH trajectories across all test cells. Notably, the model successfully captures not only the general downward trend but also subtle non-monotonic behaviors occasionally observed in lithium-ion battery aging data, such as temporary capacity regeneration (“check-up” effects after rest periods). This demonstrates the model’s strong capability to represent the complex, nonlinear degradation dynamics inherent to lithium-ion batteries.
The visual alignment between predictions and measurements is remarkably close across the entire test range, from early aging to end-of-life thresholds. The robustness of the method is evidenced by its consistent performance across multiple lithium-ion battery cells, which, despite undergoing identical test protocols, exhibit slight variations in their aging trajectories due to inherent manufacturing variability.
3.3 Quantitative Error Analysis
To provide a rigorous, quantitative assessment of the lithium-ion battery SOH estimation accuracy, we employ three key error metrics on the test set predictions:
- Mean Absolute Error (MAE): Measures the average magnitude of errors.
$$ \text{MAE} = \frac{1}{N} \sum_{i=1}^{N} | \widehat{SOH}_i – SOH_i | $$ - Root Mean Square Error (RMSE): Places a higher penalty on larger errors, reflecting estimation precision and stability.
$$ \text{RMSE} = \sqrt{ \frac{1}{N} \sum_{i=1}^{N} (\widehat{SOH}_i – SOH_i)^2 } $$ - Median Absolute Error: Provides a robust measure of central tendency that is less sensitive to outliers than MAE.
The performance of our GA-BP-ICA model for different lithium-ion battery cells is summarized in the table below. The SOH is expressed as a percentage (0-100%), and errors are reported in percentage points.
| Cell Identifier | Mean Absolute Error (MAE) % | Root Mean Square Error (RMSE) % | Median Absolute Error % |
|---|---|---|---|
| Cell #1 | 0.66 | 0.78 | 0.68 |
| Cell #2 | 0.18 | 0.32 | 0.05 |
| Cell #3 | 0.25 | 0.36 | 0.14 |
| Cell #4 | 0.08 | 0.21 | 0.01 |
The results are outstanding. The maximum MAE across all test cells is only 0.66%, and the best-performing cell achieves an MAE as low as 0.08%. The RMSE values are consistently low, with the highest being 0.78%. The extremely low median errors confirm that the vast majority of estimation points have virtually negligible error. This level of accuracy, consistently below 1% error, far exceeds the typical requirements for practical BMS applications and underscores the high fidelity of the proposed method for lithium-ion battery SOH estimation. The GA optimization effectively prevents the BP network from stagnating in poor local solutions, ensuring the model converges to a state that captures the global aging characteristics of the lithium-ion battery.
4. Discussion and Comparative Advantages
The success of our proposed GA-BP-ICA framework stems from its multi-stage design, each addressing a key challenge in lithium-ion battery SOH estimation. The ICA transformation serves as a crucial preprocessing step. By converting voltage plateaus into identifiable peaks, it distills complex electrochemical aging phenomena—such as loss of lithium inventory (LLI) and loss of active material (LAM)—into a concise set of trackable features. This provides a physically interpretable input space that is intrinsically linked to the degradation mechanics of the lithium-ion battery, as opposed to purely statistical features.
The hybrid GA-BP model addresses the core modeling challenge. While a standard BP neural network is a universal function approximator, its practical training for a complex task like lithium-ion battery SOH estimation is often hampered by random initialization. The integration of the Genetic Algorithm fundamentally improves this process. The GA performs a global, parallel search across the initial parameter space, effectively scouting for a promising starting region that is close to a high-performance optimum. This meta-optimization step confers several decisive advantages for lithium-ion battery health monitoring:
- Enhanced Convergence Speed: The BP network, starting from GA-optimized parameters, requires significantly fewer training epochs to achieve minimal error, reducing offline training time.
- Superior Solution Quality: By avoiding poor local minima commonly encountered by gradient descent alone, the hybrid model consistently finds a parameter set that yields lower final estimation error on test data.
- Improved Robustness and Repeatability: The dependence on a “lucky” random seed is eliminated. Multiple runs of the GA-BP training yield very similar high-performance models, ensuring dependable algorithm deployment for lithium-ion battery systems.
- Strong Generalization: The model demonstrates excellent performance on unseen test data from multiple lithium-ion battery cells, proving its ability to learn the underlying aging law rather than fitting noise.
Compared to other machine learning methods like standalone Support Vector Regression (SVR) or simple linear models applied to the same ICA features, the GA-BP network demonstrates superior accuracy in capturing the full nonlinear trajectory. Compared to other neural network optimizers (e.g., those using momentum or adaptive learning rates), the GA-based initialization offers a more fundamental solution to the initial condition problem. Furthermore, the entire framework is model-agnostic—it does not require an exact electrochemical model of the lithium-ion battery—making it adaptable to different cell chemistries and formats by retraining on appropriate datasets.
5. Conclusion
In this work, we have developed and validated a highly accurate and robust data-driven framework for estimating the State of Health of lithium-ion batteries. The methodology is built upon two synergistic pillars: feature extraction via Incremental Capacity Analysis and predictive modeling via a Genetic Algorithm-optimized Backpropagation Neural Network. The ICA process expertly transforms operational voltage data into salient health feature factors that physically correlate with the degradation modes of the lithium-ion battery. The GA-BP neural network then serves as a powerful, globally-optimized nonlinear regressor to establish a precise mapping from these features to the battery’s SOH.
Experimental validation on a standard public dataset confirms the exceptional performance of the proposed method. The GA-BP-ICA model achieved remarkably low estimation errors, with Mean Absolute Error consistently below 1% and often as low as a few tenths of a percent across multiple test cells. The algorithm effectively tracks the complete aging trajectory, including nonlinear capacity fade and occasional recovery effects. The integration of the Genetic Algorithm proved crucial in overcoming the convergence and local minima pitfalls of standard BP networks, resulting in a stable, fast-training, and high-precision estimator.
This work demonstrates that the fusion of interpretable feature engineering (ICA) with an intelligently optimized machine learning model (GA-BP) provides a formidable solution to the challenging problem of lithium-ion battery SOH estimation. The framework balances physical insight with computational intelligence, offering a practical, accurate, and reliable tool that can significantly enhance the capabilities of Battery Management Systems for electric vehicles and grid storage, ultimately contributing to safer and more durable lithium-ion battery applications.