Predicting when a lithium-ion battery will reach the end of its useful life has become one of the most important challenges in modern energy technology. Remaining useful life (RUL) estimates help electric vehicles avoid unexpected failures, allow energy-storage operators to schedule maintenance, and support safer operation in aerospace and industrial systems. Yet battery degradation is rarely smooth or predictable. Capacity can temporarily recover, aging patterns can change between operating stages, and the relationships among voltage, current, temperature, and charging behavior are strongly nonlinear. A new artificial intelligence framework called CA-Mamba2 is designed to address these difficulties while reducing the computational burden associated with advanced deep-learning models.
Researchers at Qingdao University and collaborating institutions developed the framework as a specialized architecture for learning how battery performance changes over time. Their work, published in ENGINEERING Chemical Engineering, combines three technologies: coordinate feature attention, the Mamba2 state-space model, and a swiGLU-gated residual network. Together, these components are intended to identify the most informative signals in battery data, retain long-term degradation patterns, and convert complex learned representations into accurate predictions of the number of cycles a battery can continue operating.
The first component, the coordinate feature attention network, or CFAN, is designed to focus on where meaningful information occurs in a battery’s degradation history. Rather than treating all measurements and time points equally, the network analyzes information along two directions: the temporal dimension and the variable dimension. This allows it to generate attention weights that emphasize important degradation stages as well as critical health indicators. For example, the model can learn that a subtle change in a charging feature may become especially valuable when it occurs during a particular phase of the battery’s aging process. The bidirectional attention mechanism helps preserve both time-dependent and feature-dependent relationships.
At the center of CA-Mamba2 is the Mamba2 backbone, a selective state-space model developed to represent long sequences efficiently. Conventional recurrent networks can struggle to retain information over long operating histories, while Transformer-based systems often require substantial computational resources because of their attention calculations. Mamba2 takes a different approach by maintaining an internal state that evolves as new data arrive. Its parameters can be selectively adjusted according to the input, allowing the model to decide which information should be retained, updated, or suppressed. For battery prediction, this provides a way to connect early signs of degradation with conditions observed much later in the life cycle.
The framework’s final major element is the swiGLU-gated residual network, or SGRN. This component applies nonlinear gated transformations to the features extracted by the attention and state-space layers. Gating enables the network to regulate the flow of information, allowing useful signals to pass through while reducing the influence of less relevant patterns. Residual connections also preserve the original representations as the data move through the network, helping prevent the loss of information and improving training stability. The resulting decoder is intended to translate highly complex degradation features into an RUL estimate without discarding important characteristics learned earlier in the model.
The researchers tested CA-Mamba2 using public battery datasets from NASA, Tongji University, and Xi’an Jiaotong University. These datasets include different battery chemistries, operating conditions, and measurement configurations, providing a demanding test of whether the model can perform beyond a single experimental setting. On NASA’s B0005 test battery, the framework was evaluated using a single input variable and at prediction starting points corresponding to 50, 70, and 90 cycles. It produced mean absolute error values between 0.0089 and 0.0098, root mean square errors from 0.0140 to 0.0172, and coefficients of determination ranging from 0.949 to 0.989. The corresponding absolute prediction errors were approximately 1.2 to 2.3 cycles.
These results exceeded the performance of several widely used comparison models, including long short-term memory networks, Transformers, PathFormer, TimeMixer, and a baseline Mamba model. The advantages became even more pronounced on the Tongji University dataset, where the researchers used 17 variables, including voltage and current statistics and features extracted from the charging process. Across prediction starting points of 200, 300, and 400 cycles, CA-Mamba2 achieved mean absolute errors of 0.0014 to 0.0016 and root mean square errors of 0.0022 to 0.0024. Its R² values reached 0.9995 to 0.9998, while absolute errors remained between 2.4 and 2.5 cycles.
One particularly important finding was the model’s stability when the prediction starting point changed. In real-world battery management, an algorithm may be asked to estimate remaining life at very different moments, depending on how much operating data are available. Some competing approaches became substantially less accurate when the starting point shifted. The absolute error of TimeMixer, for example, increased from 12.0 cycles to 50.7 cycles across the tested starting points. CA-Mamba2, by contrast, maintained almost unchanged performance, suggesting that its combination of feature attention and long-range state modeling may help it remain reliable under changing observation conditions.
The model also demonstrated strong cross-dataset performance on the XJTU battery data. In tests designed to examine generalization across batteries and degradation stages, CA-Mamba2 produced mean absolute errors between 0.0049 and 0.0083, root mean square errors from 0.0066 to 0.0111, and R² values consistently above 0.986. Computational measurements further strengthened the case for practical use. On the Tongji dataset, CA-Mamba2 required 57.848 seconds for training and 0.476 seconds for inference, compared with 423.765 seconds for Transformer training and 1.528 seconds for inference. Ablation experiments showed that both CFAN and SGRN contributed to the final gains, with their combination producing the strongest overall results. The researchers say the framework could ultimately support battery management systems in electric vehicles and stationary energy storage, where accurate, fast, and adaptable lifetime predictions are essential for safety, efficiency, and lower maintenance costs.
News Publication Date: 13-Jul-2026
Web References: https://doi.org/10.1007/s11705-026-2687-x
References: DOI: 10.1007/s11705-026-2687-x
Keywords
Lithium-ion batteries, remaining useful life, battery degradation, artificial intelligence, Mamba2, state-space models, coordinate attention, battery management systems, electric vehicles, energy storage, machine learning
