Comparative study of granular computing and RKCCA-based feature fusion for early lithium-ion battery RUL estimation Authors: K Hiroshini, Balakrishnan Devanathan, Mohammed Bou-Rabee, S Kayalvizhi Published: 2026 ABSTRACT AbstractEfficient prediction of the remaining useful life (RUL) of lithium ion batteries is essential in ensuring safety, reliability and energy efficiency of new energy-storage and mobility systems. This paper suggests a hybrid machine-learning model that combines regularized kernel canonical correlation analysis (RKCCA) with random-forest (RF) regression, hence allowing one to predict RUL accurately at early stages. In this context, the battery-degradation data can be separated in interpretable information granules by two complementary base models, Granular Model1 (GM1), which uses a small bunch of polynomies to employ degradation patterns on a low-order basis; and Granular Model2 (GM2), which utilizes deep neural networks to implement degradation movements on a higher-order platform. In the RKCCA component, fusion of nonlinear features is through the projection of both heterogeneous sensor measurements, such as voltage, current, and capacity, to an assumed common space in the latent features. This is a transformation that improves the correlation between features and also alleviates effects of noise and the dimensionality in the data. The resulted fused latent representations are used as inputs to the RF regression model to give the final RUL estimate. The application of the experimental analysis to the dataset of MIT Battery Degradation (A123 LFP/Graphite cells) reveals that the suggested RKCCA + RF model significantly exceeds GM1 and GM2. The model outputs a mean relative error of 3.45 cycles, root-mean-square error of 4.75 cycles and coefficient of determination () of 0.9995. Overall, the offered results are the testimony to the fact that the model is a high-ability framework that is able to model nonlinear dynamics of degradation and provide high-accuracy predictions, and that can be implemented effectively in the sphere of data-driven prognostics and predictive maintenance and real-time battery-management systems.Keywords: Remaining useful life, Lithium-ion battery, Polynomial regression, Deep neural networks, Granular computing, RKCCA, Random forestSubject terms: Energy science and technology, Engineering, Materials science, Mathematics and computing Introduction Lithium-ion batteries find extensive applications that include smart phones, electric car and renewable energy systems on the grid. The importance of these batteries has emerged as a vital element of contemporary technological infrastructure, causing a growing interest in enhanced levels of safety, dependability, and high-performance longevity of operation distance of a battery device1–3. As a result, the accurate tracking of the battery health has become a primary research area. Proper estimation of RUL which refers to the duration of time that a battery can be used with specified conditions under which the battery functions well under conditions given to them is one of the key factors in the battery health management4. Predicting RUL reliably makes it possible to perform proactive maintenance and thus minimise operational costs, increase system safety, and optimise generally the performance of a system5. Current prediction methods in RUL can broadly be divided into physics based and data-driven models of RUL prediction. Interpretable physics-based models, such as electrochemical models and equivalent circuit models, were used to gain an understanding of the processes of degradation, including Solid Electrolyte Interphase (SEI) growth and active lithium loss, considered separately, in detail, and outside a certain context6. Nevertheless, such models can be difficult to identify parameters and they can be inapplicable to different chemistries and operating backgrounds. Electrochemical model As an example, Lyu et al.7 created an electrochemical model which used particle filtering of lead-acid batteries, though the algorithm was computationally expensive and very sensitive to sensor accuracy, as described in. By contrast, data-driven methods are not based upon clear physical postulations, but learn degradation patterns directly based on operational data. Predictive interval and uncertainty quantification offered by regression-based approaches including Gaussian Process Regression (GPR) is useful in decision-making under uncertainty8,9. Support Vector Machines (SVMs) operate on the basis of kernel functions to learn nonlinear degradation properties although they may need a lot of hyperparameter optimization and do not have an inherent uncertainty assessment mechanism10. Recent neural network architectures such as Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs) show great promise in identifying high-level temporal information in battery data11,12. However, these models are generally bulky and they can overfit when dealing with early cycle data or small samples. In addition, they cannot be used in critical safety-based applications because of their low interpretability. To overcome the weaknesses of black-box learning models, a recent study has introduced Granular Computing (GrC) that has been found useful in solving problems that struggle with such models13. GrC illustrates knowledge through the application of information granules which are interval-based representation and express uncertainty and relevance explicitly in predictions. Granular models offer measures of coverage and specificity to decision support based on uncertainty in predictive maintenance situations and procedural support, which is uncertainty conscious, is present in the model of granularity, reduction in predictive maintenance decision-making models, therefore, helps to achieve decision support in predictive maintenance problems14. As an example, Zhang et al.15 suggested a hybrid prognostic model of a combination of Random Forest (RF) and Artificial Bee Colony optimization, and General Regression Neural Network. Machine learning-based hybrid models have also been studied in addition to stochastic and random process-based ones as far as RUL prediction is concerned16,17. Wiener process models have been shown to be rather successful in the process of capturing the uncertainty and unpredictability in battery aging curves. Such models define degradation as a drift and diffusion continuous time stochastic process degradation model or degradation model partitions. Such models, though useful, usually require a lot of parameter optimization and might not be able to scale well to a wide variety of data. Recently, signal-processing and decomposition-based hybrid pipelines are also investigated to consider nonlinear and non-stationary battery degradation signals to solve them18,19. Multi-stage optimization and attention techniques lead to advanced deep hybrid frameworks such as VMD-SSA-PatchTST and spatial-attention-enhanced-TLSTM-CNN systems that establish state-of-the-art accuracy with attention mechanisms and without requiring additional training stages to achieve ultra-high accuracy on par with advanced end-to-end machine learning frameworks up to now20,21. The models however are usually computationally intensive, sensitive to hyperparameter choices and not easy to interpret restricting their use in real time battery management systems. In the same way, the LSTM-Transformer hybrid architectures and the use of metaheuristic-based optimization systems have a high predictive ability with the drawback of not focusing on interpretability and uncertainty explicit representation either in their design or algorithmic implementation22–24. Recent review literature reveal that although there has been major progress on physics-based, data-driven and hybrid models, interpretable and uncertainty aware feature-fusion methods are underdeveloped, especially when it comes to early cycle RUL prediction. Although multi-model and decision-level fusion strategies with reinforcement learning-based fusion networks have demonstrated promising processes26, most of them are based on black-box architecture and are not transparent. Motivated by these limitations, this study proposes an RKCCA-based feature-fusion framework integrated with granular computing and Random Forest regression for early-stage RUL prediction. The proposed method performs nonlinear fusion of heterogeneous battery signals (voltage, current, and capacity) using Regularized Kernel Canonical Correlation Analysis (RKCCA), projecting them into a shared low-dimensional latent space where nonlinear dependencies are preserved and noise effects are mitigated. Random Forest regression is then employed to model degradation dynamics while maintaining robustness and interpretability. Granular computing further enables explicit uncertainty representation through interval-based information granules. The proposed framework adopts a two-objective optimization strategy that balances predictive accuracy and model interpretability. Unlike many deep hybrid models that achieve high benchmark performance but exhibit limited robustness in real-world deployment, the present approach emphasizes computational efficiency and scalability for early-cycle prediction under limited data conditions. This design supports sub-second inference, making it suitable for real-time battery management systems in electric vehicles, industrial Internet of Things (IoT) environments, and large-scale renewable energy storage applications (Table 1). Table 1Comparison of RUL prediction methodologies.MethodPrediction typeUncertaintyAccuracyElectrochemical8NumericalX√GPR9ProbabilisticSVM10Point estimateXLSTM11Point estimateXGranular model14Interval-based Comparison of RUL prediction methodologies. Proposed RUL prediction methodology It is suggested in this work as shown in Fig. 1, A hybrid perspective of accurate, reliable as well as explanatory advance-prediction of the RUL of the lithium-ion batteries by means of RKCCA combination and RF regression. This framework relies on the data of battery degradation at the early stages (like voltage, current, capacity, temperature, and cycling data) that result in the estimate of RUL at a timely manner, although at reduced cost in an experiment. Normalization, temporal alignment, and noise reduction are first used as pre-processing to ensure consistency and nature of values across operating cycles of the raw sensor values.RKCCA is then used to incorporate multivariate measurements into a common latent feature space to best explain complex nonlinear relationships between heterogeneous sensor signals, and to regularize overfitting. The latent features obtained give the compact and informative representations of battery health evolution and on which further granular analysis is based. There are two complementary granular computing models that are applied to achieve high standards of interpretability and reliability: GM1 which is based on the use of a deep neural network to model nonlinear and temporal higher order degradation trends based on a regular regression model, and GM2 which is based on the use of a deep neural network to model nonlinear and temporal high-order degradation trends applied to a regular regression model. The results of these models are converted into information granules (intervals based) and the quality of the granularity is measured using coverage and specificity to guarantee a well balanced and meaningful reflection of uncertainty. Lastly, a Random Forest regression model is trained on the RKCCA-derived latent features and granular representations to predict RUL using its insensitivity to overfitting and high-dimensional feature space nonlinear interactions. The presented framework generates point RUL estimates and interval-based uncertainty bounds and its performance is measured based on MAE, RMSE, the coefficient of determination (), coverage, specificity, and the composite granular quality measure (Q). Block diagram of the proposed RKCCA–RF-based framework for early lithium-ion battery RUL prediction. Data collection and preprocessing This paper uses the open-access MIT Battery Degradation Dataset that includes bicycling data of commercial A123 lithium iron phosphate (LFP)/graphite cells under accelerated degradation conditions. The batteries were rotated until end-of-life, which was the stage when the discharge capacity reduced to 80 percent of nominal capacity. The data set has high-resolution data that are measured in and across charge discharge cycles in terms of voltage, current, charge discharge capacity, temperature, internal resistance, and charge time, respectively, and thus capture all the electrochemical and thermal dynamics under realistic operating conditions. To allow early-stage prediction of the RUL, the analysis will be restricted to early-cycle degradation data, which permits timely prognostics with a relatively low cost of experiment and low data needs.Representative discharge capacity degradation trajectories are shown in Fig. 2a, demonstrating a clear and monotonic capacity fade suitable for RUL modeling. In addition, representative voltage and current profiles during early charging cycles shown in Fig. 2b,c have a stable and repeatable behavior, which encourages the construction of health indicators using charging-phase measurements which tend to be more consistent than the discharge profiles. Overview of data collection and preprocessing workflow and results. A feature selection strategy that was implemented was a sliding window, which was used to determine the most informative parts of a cycle. The quantification of the relationship between partial-cycle degradation fragments and final cycle life was done using Pearson correlation analysis. Cycles between 200 to 500 cycles resulted in the most consistent and best correlation with the end-of-life capacity, as shown in Fig. 1 below Fig. 2d, thus being a most suitable choice to analyse further. This correlation-based methodology guarantees that maintained pieces of data still have the greatest amount of the degradation worth and have the capacity to be predictive at an early point. Based on the estimated cycle windows, nine indirect health indicators (HIs) were calculated mainly based on measures at the charging step, reproducing the complementary degradation properties in the form of voltage development, current activity, capacity drift, thermal reaction and defining resistance dynamics. Figure 2e shows the normalized evolution of representative health indicators across early cycles which are smooth and degradation sensitive trends which can be modeled accordingly. Moreover, the Pearson correlation table presented in Fig. 2f below reveals sophisticated interdependencies between the obtained HIs and how they correlate with the cycle life, which inspires the feature fusion algorithms to use include RKCCA, PCA. All the extracted health indicators underwent preprocessing before model input was done to strengthen its robustness and consistency. This preprocessing involved signal smoothing to reduce the effects of high-frequency signals, minmax normalization to make comparison across indicators comparable, and political modeling to generate stable degradation curves. The Fig. 2g below can be used to present an illustrative comparison between the raw and preprocessed health indicator indicator signals which illustrates the success of noise reduction and stabilization. These preprocessing tasks guarantee a good condition of the resultant features to be used in fusion of features, granular modelling and regression to predict the RUL. Lastly, the suggested preprocessing plan was implemented on the MIT battery degradation dataset, which was composed of 123 lithium-ion cells that were subjected to various charge/discharge cycles allowing to evaluate the suggested framework systematically. Granular model 1: polynomial regression GM1 forms the foundational layer of the predictive framework by providing a transparent, structurally simple, early-cycle RUL estimation approach. It relies solely on simple numerical relationships from the initial operating behavior of the battery and, therefore, is well-suited for establishing a clear baseline before any more sophisticated elements of modeling are introduced. Within the hierarchy, GM1 provides a point of interpretability that allows subsequent granular models to be compared uniquely and meaningfully. GM1 thus stands as the simplest and most interpretable benchmark.GM2 is a natural extension of GM1 through the inclusion of DNNs to elaborate on nonlinear, advanced feature abstractions and GPR for uncertainty quantification. The underpinning idea behind this framework extends to processing each extracted HI within its building block, hence a group of modular neural sub-networks, and processes the raw extracted HI. The predictions from the modular sub-networks are then combined using a GPR layer to produce a probabilistic output (mean and variance) that, in turn, will be transformed according to the principle of justifiable granularity into an interval-based granular output with intelligent trade-offs between coverage and specificity for the prediction. Granular model 2: deep neural networks GM2 starts with the representational chaining offered by data-driven, learning-based mechanisms for the relevant modeling of nonlinear degradation behavior. GM2 thus extends beyond the purely numerical baseline toward a deep-learning structural architecture with the express purpose of capturing more complex temporal patterns in the health indicator series of the batteries. It forms the intermediate level in the hierarchy, with some loss of predictive flexibility compared to the latter and without compromising the architecture’s systematic positioning for granular transformation. Entropy-based feature extraction Entropy-based features are employed to quantify the uncertainty, irregularity, and complexity inherent in lithium-ion battery degradation signals. These features provide a compact statistical representation of nonlinear aging dynamics that may not be fully captured through conventional time-domain indicators. Given a discrete time-series signal extracted from voltage, current, or capacity measurements over early charge–discharge cycles, the probability distribution of the signal amplitudes is estimated using histogram binning or kernel density estimation. The Shannon entropy, which measures the average information content of the signal, is defined as: where M denotes the number of discrete bins and represents the probability of samples falling within the i-th bin. To further characterize temporal complexity and self-similarity in degradation evolution, Sample Entropy (SampEn) is adopted. For an embedding dimension m and tolerance r, SampEn is expressed as: where A denotes the number of matching template vector pairs of length m+1 within tolerance r, and B denotes the number of matching template vector pairs of length m. These entropy-based descriptors capture the degree of disorder and stochasticity in early-cycle battery behavior, offering robust and interpretable indicators that complement conventional health indicators for subsequent feature fusion and RUL prediction. Principal component analysis Principal Component Analysis (PCA) is applied to reduce feature dimensionality and mitigate multicollinearity among the extracted entropy-based and health indicator features. PCA enables the transformation of correlated input variables into an orthogonal subspace while preserving the maximum variance in the data, thereby enhancing numerical stability and model generalization. Let denote the zero-mean feature matrix consisting of n samples and p extracted indicators. The covariance matrix of is computed as: PCA seeks projection vectors that maximize the variance of the projected data subject to orthogonality constraints: This optimization leads to the eigenvalue problem: where denotes the k-th largest eigenvalue. The principal component scores are then obtained as: where contains the eigenvectors corresponding to the K dominant eigenvalues. The resulting low-dimensional representation retains the most informative degradation patterns while suppressing noise and redundant information, thereby facilitating efficient and robust feature fusion in subsequent RKCCA-based modeling. Regularized kernel canonical correlation analysis The RKCCA algorithm is used on nonlinear fusion of features used in heterogeneous sensor domains such as voltage, current and capacity. The general layout of the RKCCA framework that is to be applied in this case is depicted in Fig. 3. In contrast to traditional KCCA, where it is necessary to compute all eigenvalues and eigenvectors of large covariance matrices, RKCCA uses a low-tech and effective regularization method. The method is crucial in decreasing the dimensional complexity and retaining the nonlinear interactions between features that are most useful in early-cycle battery prognostics. With the traditional KCCA, two generalized eigenvalue problems need to be solved, which are resource-intensive in the case of large-scale data. Due to this reason, RKCCA does a single iteration of the power method and only the most dominant eigenvalue and eigenvectors are obtained. This can be stated, assuming a matrix D is diagonalizable and nonzero initial value b0 is provided, the eigenvector of the largest eigenvalue of D can be calculated with the following recursion. : General architecture of the Regularized Kernel Canonical Correlation Analysis (RKCCA) framework for nonlinear feature fusion across heterogeneous battery sensor domains. Through this iterative approach, the relationship between the canonical weight vectors α and v can be derived as follows: where K and Y denote the kernel and feature matrices of the input domains, and η is a small regularization coefficient ensuring numerical stability. The iterative update continues until convergence between α and v is achieved. After obtaining the optimal α and v, the prediction for new data samples is computed. The kernel matrix of the new inputs is defined as: To ensure centered kernel features, is normalized as: Let represent the score matrix of , which can be expressed as: Considering as the corresponding score of , the following relation holds: and the prediction for is finally computed as: where denotes the Moore–Penrose pseudoinverse of v. This iterative RKCCA formulation avoids full eigenvalue decomposition of large covariance matrices. Here, the maximal eigenvalue and the associated eigenvectors are efficiently estimated by performing a number of matrix multiplications. It reduces the computational burden at the cost of only slight accuracy loss. Regarding this, RKCCA presents a scalable and highly effective application in feature extraction for large multisensor data sets, which is explicitly significant for battery degradation modeling and RUL prediction in high dimensionality. Random forest regression and RUL estimation The representations of features in the latent state achieved by the RKCCA transformation are subsequently used as inputs to a Random Forest (RF) regressor for estimating the RUL. Random Forest is an ensemble learning method that constructs multiple decision trees using bootstrap aggregation and random feature selection, improving robustness, reducing variance, and enhancing generalization for nonlinear regression problems. Let denote the training dataset, where represents the RKCCA-derived latent feature vector and denotes the corresponding true RUL value. The RF model consists of regression trees that make up the RF model were each trained using a bootstrap sample taken from . The mean squared error (MSE) is minimized at each split to determine the best partition: where and denote the left and right child nodes, respectively, and is the average RUL value of samples inside node , and and represent the left and right child nodes, respectively. The average of each tree’s prediction yields the RF regression output for a new input feature vector : The RF model is expanded to produce prediction intervals by utilizing the distribution of outputs from individual trees in order to further quantify prediction uncertainty for interval-based evaluation. Each tree generates a prediction given an input feature vector . We create an empirical distribution rather than depending only on the mean estimate: The prediction interval is then obtained using quantiles of this distribution: where the interval’s confidence level is controlled by . This method offers a non-parametric mechanism for uncertainty estimation without making any assumptions about a predetermined distribution and is compatible with Quantile Regression Forests. The RF model can now generate interval-based predictions thanks to this extension. These predictions are then assessed using Granular Computing metrics, such as Coverage, Specificity, and Quality (Q). Granular computing-based evaluation metrics GrC metrics are used to evaluate the prediction intervals derived from the RF model. These metrics evaluate the prediction intervals’ compactness and dependability. Coverage The percentage of actual RUL values that fall inside the anticipated range is known as coverage: where is the indicator function, which equals 1 if the condition is satisfied and 0 otherwise. Specificity Specificity evaluates the compactness (sharpness) of the prediction intervals: where narrower intervals result in higher specificity values. Quality (Q) The Quality metric balances coverage and specificity: where controls the trade-off between reliability and compactness. Overview of the RKCCA+RF hybrid framework This paper contrasts a hybrid machine-learning approach that estimates RKCCA-based RF regression against Grc-based models for early-cycle RUL prediction. The fundamental idea is based on using RKCCA as a feature engineering tool for deriving a compact but highly informative latent feature space from complex multi-view sensor data. The transformed features are analyzed from a comparative point of view based on reasoned differences in interpretability, predictive capacity, and computational cost across the GM1 and GM2 models relative to the RKCCA + RF model. While GrC models mainly focus on an interpretable, interval-based prediction that naturally accommodates uncertainty for risk-aware decision-making, the RKCCA + RF approach attempts to uncover deep, nonlinear relationships among multi-source battery voltage, current, and capacity in predicting cycle life. This is the granularity concerning the impact of statistical granulation on early RUL estimates in performance, scalability, and interpretability, against which the non-linear feature fusion is compared. Second, GrC has modeled RUL as an interval-based prediction, hence capturing fundamental uncertainty quantification, which is crucial for risk-aware maintenance decisions. The subsequent application of RKCCA herein addresses the efficiency with which deeper, more complex relationships can be sought among multi-view battery signals in attaining eventual cycle life. Goodness-of-fit is more straightforward in this respect since it automatically identifies those patterns in data that are most predictive of the battery life without substantial input to manual feature engineering. Indeed, as part of KCCA, RKCCA is the computationally efficient version of KCCA, which has enormous POS devices practical for near-real-time applications. The methodology across cycles begins with sequence preparation and seamless data integration into the RKCCA procedure. Each sequence is analyzed for its relevance concerning battery lifespan, and only the sequences that can be validly considered in relation to a strong predictive relationship for the output variables are validated. The RKCCA transformation does indeed hold the multi-view features in a shared latent space that enhances stability and high clarity as to the degradation patterns clearly characterized by the downstream granular models. Results and discussion The experiment was done on the MIT Battery Degradation Dataset of A123 LFP/Graphite cells under charge-discharge 100 cycles. The computations of each cycle, the HIS mean voltage, maximum current, and discharge-capacity mean, were normalized and taken as predictive features. The RUL as a dependent variable. The data were divided into training and testing sets in the proportion of 80:20 and tested via six conventional indicators: MAE, RMSE, , Cov, Sp and Composite Accuracy(Q). RKCCA + RF is a dimensionality reduction and feature fusion ally scheme, which makes disparate input features embedded in the unified latent space. These different views include raw multi-view data discharge capacity (Q(V)), voltage signal versus time (V(t)), incremental capacity (IC), resistance indicator (V/I(t)), and interpolated current (ICC), and these views are transformed to obtain a highly parsimonious and insightful latent space. The latter step does play a very significant role in maximizing the stability of the model at this point, refining the modest model prediction to explain the connection between features and de-noise. Granular model 1 performance analysis As it is presented in Figs. 4a–c respectively, three GM1 structures were tested and compared, namely: (i) a simple linear regression model, (ii) a tuned normalizing-polynomial regression model, and (iii) a deep neural network (DNN) model, to evaluate the predictive ability of the granular computing framework on the basis of the available predictor variables. Use of the basic GM1 regression, which employed ordinary least squares regression, was used as a starting point, and provided a mean absolute error (MAE) of 42.41 and a root mean square error (RMSE) of 50.40. These findings show that it has a poor ability to capture nonlinear degradation processes of lithium-ion battery data. By contrast, the tuned GM1 model showed a great improvement with an MAE of 22.86 and RMSE of 28.83. The reason why this performance improvement was obtained can be explained by the feature normalization and second-order Polynomial expansion that boosted the capacity of the model to explain the nonlinear relationships between various degradation models with limited computational cost. The GM1-DNN model (two hidden layers with dropout regularization of 128 and 64 neurons) had a MAE of 23.34 and a RMSE of 28.77. It had a similar predictive performance with that of the tuned GM1 model. But at this low cost, more computational complexity and less interpretability occurred. In general, the tuned GM1 model has become the most appropriate model to use in case of early-cycle RUL estimation and it provides a good trade-off between predictive performance, computation speed and model interpretability as it can be seen in Table 2. GM1 model: actual vs predicted RUL prediction graph. Table 2Overview of the three GM1 models.ModelAlgorithm typeTechniques usedMAE (cycle)RMSE (cycle)GM1 (base)Linear regressionSimple linear mapping42.4150.40Tuned GM1Polynomial regression (degree = 2) + NormalizationStandardScaler + Polynomial features22.8628.83GM1 (DNN)Deep neural networkDense layers + Dropout + early stopping23.3428.77 Overview of the three GM1 models. Granular model2 performance and feature behavior analysis Deep learning model evaluation Task The GM2 configuration used a DNN version that had two hidden layers consisting of 128 and 64 neurons, respectively, and dropout regularization. The model was then trained with Adam optimizer that has adaptive learning rate and an early-stopping mechanism was implemented when training to prevent overfitting. Figure 5 shows that after the application of a Gaussian kernel to smooth the data exponential-kernel graph, the actual and predicted values of RUL are compared to each other in Fig. 3 after the application of the technique. This demonstrates that the degradation curve that was predicted was behaving as per what is expected of the actual trend in terms of RUL especially in mid-cycle regions but started to shift by a little bit in the terminal cycles as expected of the actual degradation curve since the density of the data reduced. On a quantitative level, it was doing better than the DNN-based solution, with MAE = 23.67 cycles and RMSE = 30.82 cycles. Additional validity of adequate performance was provided, in fact, establishing improved nonlinear learning in the GM2 over the GM1 baseline of polynomials. Loss oscillation is also highly insignificant and convergent which shows a promising stability effect. On the whole, the DNN approach demonstrates a better flexibility and sensitivity of data than any other regression techniques to capture the relationship between CC and CD as well as the course of degradation in a nonlinear manner. Exponential-kernel graph—GM2 deep learning (actual vs predicted RUL). Computational complexity analysis Computationally the GM1 model can be solved with a linear regression whose closed-form solution has computational complexity of a linear regression of complexity of the order of:, n the number of samples and d the number of input features. This renders GM1 to be computationally efficient during training and inference. There is an increment of tuned GM1 model, which adds new smoking steps to preprocess features, such as normalization of a feature and expansion of features using polynomials that combine to enhance the fundamental feature dimension (d' > d). Consequently, the computational complexity is of the order of the square of the dimensions of the d-array, which is a smaller order of magnitude, resulting in a moderate increase in the cost of computation compared to before, but still an efficient train because of a linear structure. Conversely, GM1-DNN model is much more complex than the other as it uses a multi-layers neural network that is developed through iterative backpropagation. This complexity of the training can be stated at the scale of: , where by E is the number of training epochs, and n h is the quantity of hidden units. This increases the computational needs in training as opposed to GM1-based models. Nevertheless, this added complexity can be linked to the training stage mostly. Adoption Inference Under inference, the GM1-DNN model only computes forward-pass operations, which comprise of matrix multiplications and nonlinear activations, and is uniformly computational. Thus, the GM1-DNN model does not only increase the complexity of the training process, but it should also be used in real-time prediction applications because of the low inference cost. The enhanced predictive accuracy, in such a way, comes to the price of the increased training complexity, which is a trade-off between the accuracy and the computational efficiency. Feature behavior and contribution analysis In this respect, a thorough analysis of the extracted HIs has been performed to further investigate interpretability and feature relevance with respect to the GM2 model advocated within this work. Each is linked to the different electrochemical or operational lithium-ion battery mechanisms of degradation. Therefore, the evolution of these indicators over time concerning the cycle index first suggests the promising nature of the sensitive features with respect to the early aging dynamics. Figure 6a shows the trend of HI (duration) with respect to the cycle index. This indicator reflects the duration of discharge per cycle, so it can be valuable for monitoring the rate of loss in usable capacity. A drastic fall within the first few cycles is representative of rapid degradation due to the decomposition of the electrolyte and the formation of the solid-electrolyte interface; hence, this would be followed by a slower loss as diffusion-controlled capacity loss dominates the later stages. Because its very early downward slope within the curve suggests that this HI would become a responsive marker to suggest the onset of degradation, however, Fig. 6b presents the profile of increasing HI (max voltage). The maximum voltage across charging cycles gradually increases as the internal resistance builds up and finally dominates the influence of electrode polarization. Therefore, the typical recovery and decay in voltage characteristic of LFP chemistry mark this point in the process where reversible aging gives way to irreversible degradation. In this case, voltage-based metrics become decisive in tracking the dynamics and trends of remaining energy capacity during the later stages. Behavior of selected health indicators (HIs) across charge–discharge cycles. The extracted HIs have a direct impact on the precision of RUL prediction and are essential in capturing the degradation behavior of lithium-ion batteries. Each HI reflects a different facet of battery dynamics, including capacity variation, voltage evolution, and current response, all of which are intimately linked to underlying electrochemical aging mechanisms. A more thorough depiction of battery health is made possible by the use of several heterogeneous HIs, which enable the model to capture both linear and nonlinear degradation patterns. Additionally, the drop-one-HI analysis, in which the elimination of individual indicators results in discernible changes in prediction performance, validates the efficacy of the chosen HIs. This shows that the extracted HIs are crucial for reliable and accurate RUL estimation and provide complementary information. Drop-one-HI sensitivity evaluation In order to measure the value of each indicator, a drop-one-HI analysis was conducted. In this experiment, HI were removed sequentially in the input feature set and GM2 deep neural network was re-trained to evaluate the consequence on model performance. Figure 7 represents five major evaluation measure standards, namely RMSE, MAE, Coverage, Specificity and Q-metric, which the retraining process with the considered option of removing each HI. Thus, HI will be best represented in capturing the degradation by the fact that, as Fig. 7 drop indicates, disabling HI (current integral) to train will double both RMSE and MAE, since this component models terminal-phase behaviour. Figure 8 shows the results within the testing phase, after which a similar tendency was found in sensitivity. The emission of the HI filtrates HI and HI resulting in a gradual rise in RMSE and a decline in Q indicating that the success in this prediction mainly depends on voltage and current based measures. Nevertheless, most cases of HI do not exhibit a detergent change in terms of both the Coverage and Specificity thus there is no single indicator that triggers the overall model generalization that is the combined effect of the two in terms of each other in the case of HI. GM2: training metrics—intuitive distribution of evaluation metrics by dropping one HI. GM2: testing metrics—intuitive distribution of evaluation metrics by dropping one HI. Quantitative impact of individual health indicators Table 3 summarizes the quantitative impact of excluding each health indicator (HI) on the predictive accuracy of the GM2 model. The most significant increases in MAE and RMSE occur when the integral features of current and voltage (HI and HI) are removed, indicating that accumulated charge and energy-related information play a critical role in long-term degradation forecasting. In contrast, excluding the average current and voltage indicators (HI–HI) results in only minor performance degradation. This suggests that these features act as secondary descriptors, whose information content is largely captured by other dominant indicators within the early-cycle predictive framework. Table 3Quantitative metrics for each dropped HI.HI droppedMAERMSECoverage (%)Specificity (%)Q metricHI duration23.1927.423268-0.04HI voltage integral25.6632.7140600.20HI current integral27.4833.182872-0.16HI avg voltage24.6231.1140600.20HI avg current22.6828.4736640.08HI max voltage21.9628.2636640.08 Quantitative metrics for each dropped HI. Interpretation and discussion The findings illustrated in Fig. 8 and in Table 4 suggest that GM2 predictive capability lies in many combinations of HIs. Time-based indicators, for instance, HI , are sensitive to early aging stages and track rapid declines in performance very effectively. For medium- to long-term forecast accuracy, the integral features HI (Voltage Integral) and HI (Current Integral)are developing since they reflect the cumulative interaction of the electrochemical processes over cycling. On the other hand, HI (Maximum Voltage) is more sensitive to end-of-life phenomena related to internal resistance and polarization. Overall, these findings reinforce that GM2 performance rests on the synergistic interaction among many HIs rather than on any simple substrate of one or several features. This appropriately balanced contribution from the inputs enables the model to properly capture both early and late degradation behavior for accurate and consistent RUL predictions. Table 4GM2 drop-one-HI testing results.Dropped HIRMSEMAECoverageSpecificityQHI1133.5452118.02541.0− 0.2491− 0.2491HI2134.2604121.19521.0− 0.2493− 0.2493HI3130.3092117.12581.0− 0.2239− 0.2239HI4134.2812118.22181.0− 0.2567− 0.2567HI5136.0312120.64131.0− 0.2805− 0.2805HI6133.0111118.38771.0− 0.2719− 0.2719HI7136.0312120.64131.0− 0.2805− 0.2805HI8136.0312120.64131.0− 0.2805− 0.2805HI9134.4211118.39571.0− 0.2688− 0.2688 GM2 drop-one-HI testing results. Interpretation and discussion: EF, PCA, and RF-based models The analysis of the gradual effects of the feature abstraction and regression strategies before the fusion with RKCCA is done by the performance of Entropy-Based Features (EF), Principal Component Analysis (PCA) and Random Forest (RF) models in terms of actual and predicted RUL plots and relative error measures such as MAE, RMSE and . Figure 9a the real vs the predicted RUL values given by the model built using the EF-based model are represented against a test set in Fig. 1. As it can be seen, despite the global tendency of adherence to the degradation path, there are also strong deviations of an ideal regression line to all the range of RUL values, in particular, to high RUL values. The PCA-based regression results are shown in Fig. 9b. Compared to EF, PCA improves alignment between predicted and actual RUL values by mitigating feature redundancy and projecting the input space onto dominant variance directions. However, the non-optimal fit is still noticeable, which indicates the intrinsic weakness of linear projections to capture nonlinear interactions of higher order involving the indicators of battery degradation. The prediction results using RF are shown in Fig. 9c. It is evident that there is an improvement and the predicted RUL values are concentrated close to the 45-degree reference line. This confirms that ensemble learning is effective in the modeling of nonlinear relations and variance reduction by the aggregation of trees. Figures 10a (mae), b (rmse) and c (r2) provide the summary of the comparative quantitative performance of EF, PCA, and RF models. According to MAE comparison, RF has the lowest prediction error followed by EF and PCA. In the RMSE comparison, as well a similar trend is noted with RF significantly reducing large deviation errors, which is a measure of enhanced robustness. Moreover, the comparison of the shows some noticeable performance hierarchy as RF has contributed the largest percentage of the variance in the battery cycle life, followed by PCA and EF which have consecutively lower goodness of fit. In order to measure the generalization behavior, we compare training and testing MAE of the three models in Fig. 11. The difference between the training and the testing error of RF is small indicating its high generalization ability with fewer overfits. Conversely, EF and PCA gaps are larger, which is the sign of deteriorating predictive stability. Differently put, it can be concluded that an improvement occurs when the models switch at EF to PCA to RF which in turn leads to the conclusion that entropy based descriptors, PCA and Random Forest regression use degradation behavior. Actual vs. predicted remaining useful life on the test set for EF, PCA, and RF-based models. Comparative performance evaluation of EF, PCA, and RF-based models using MAE, RMSE, and metrics. Training vs. testing MAE comparison for EF, PCA, and RF-based models. Overall comparison Table 5 gives an overall comparative analysis of the proposed RKCCA + RF model versus traditional regression models, deep learning baselines, early-feature machine learning models, and health-indicator (HI) ablation variants, the point-prediction accuracy and granular reliability measures. Baseline granular models GM1 (Polynomial Regression) and GM2 (Deep Neural Network) display comparatively high prediction errors where the RMSE value of more than 28 cycles demonstrates that the two models are not effective when the application of granular models is made without a strong fusion of features. The substantial decrease in RMSE values in early feature-based models like PCA + RF and EF + RF compared to GM1 and GM2 indicate that machine learning regression has a significant edge; nevertheless, their intermediate RMSE values (16–19 cycles) and lower values of less than 0.7 indicate that the first is limited in determining the complex nonlinear degradation dynamics of lithium-ion batteries. The HI-drop analysis additionally indicates that de-setting individual health indicators in GM2 results in unstable granular performance, with a number of settings giving low or negative Q values, which indicates low specificity or uncertainty poor calibration and the shortcomings of isolated or manually selected indicators. Conversely, the proposed RKCCA + RF model demonstrates significantly low comparative methods, with an RMSE of 4.75 cycles, an MAE of 3.45 cycles, and a high of 0.9995, and at the same time, good granular reliability coverage of 0.9375, specificity of 0.9115 with the high Q value of 0.8545. It has been explained by a superior performance due to the fact that heterogeneous sensor data is effectively fused by the use of RKCCA which isolates a high degree of correlated latent features that are less susceptible to noise, and the strength of Random Forest regression to model nonlinear interactions. On the whole, the findings prove that the proposed framework does not only offer meaningfully better prediction accuracy but also effective and reliable uncertainty quantification, which makes it a reliable and feasible tool in predicting RUL in the early stage and a convenient way of managing batteries (Fig. 12). Table 5Comprehensive performance comparison of RUL prediction models.ModelRMSEMAECoverageSpecificityQGM1 (polynomial regression)28.8322.86––––GM2 (deep neural network)30.8223.67––––PCA + RF (early)18.6111.65–––0.6055EF + RF (early)16.287.79–––0.6983HI-drop (GM2) – HI Duration27.4223.190.320.68-0.04–HI-drop (GM2) – HI Voltage integral32.7125.660.400.600.20–HI-drop (GM2) – HI Current integral33.1827.480.280.72-0.16–HI-drop (GM2) – HI Avg voltage31.1124.620.400.600.20–HI-drop (GM2) – HI Avg current28.4722.680.360.640.08–HI-drop (GM2) – HI Max voltage28.2621.960.360.640.08–RKCCA + RF (proposed)4.753.450.93750.91150.85450.9995 Comprehensive performance comparison of RUL prediction models. Performance comparison of battery RUL prediction models. Conclusion and future work The current paper provides a comparative study of GM1, GM2 and the suggested RKCCA + RF model to predict the RUL of lithium-ion batteries. These results show that the proposed model performs better than the baseline methods, recording an MAE of 3.45 cycles, RMSE of 4.75 cycles and coefficient of determination () equal to 0.9995, which is a highly accurate predictor. The RKCCA + RF framework benefit consists of the fact that it is capable of successfully merging heterogeneous degradation signals via RKCCA-based nonlinear feature extraction and random forest regression is employed to obtain complex nonlinear relationships with the possible threat of overfitting minimized. Besides better accuracy, the proposed methodology will characterize uncertainty based on granular measures, thus improving the interpretability. Despite the fact that the assessment was done on one battery data, the findings indicate that hybrid kernel-ensemble and granular computational systems have some promising potentials in early-stage battery health prognostics. Nonetheless, additional experimentation on variety of battery chemistries, operating environments, as well as real-world data is needed to establish generalization. Future work will be done to put the framework to various battery chemistries and various working conditions, physics informed constraints, and explainable artificial intelligence methods to enhance robustness, transparency, and possible applicability in battery management systems. Author contributions H.K., B.D., M.B., and S.K. wrote the main manuscript text, and H.K., M.B., and B.D. prepared Figs. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 and 11. All authors reviewed the manuscript. Funding Open access funding provided by Amrita Vishwa Vidyapeetham. The authors received no external funding for this research. Data availability The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request.