An automated method for neural architecture design in lithium-ion battery state of health estimation

Lithium-ion batteries are essential to modern energy systems, and accurately estimating their State of Health (SOH) is vital for ensuring their reliability and lifespan. Data-driven based methods, especially those utilizing deep learning, are emerging as a prevalent way for SOH estimation. Nevertheless, these neural networks are typically designed using a trial-and-error paradigm that is guided by experts. In order to shift this paradigm from expert-driven to automated design, this paper proposes an automated method for constructing neural architecture for SOH estimation. Specifically, the charging voltages are first encoded into an image to enrich the representative capability. Then, we formulate the architectural design of SOH estimation as a multi-objective search problem and simultaneously consider accuracy and computational complexity. An evolutionary method is then employed to solve this problem, yielding a Pareto set that offers accuracy-best, low complexity, and balanced trade-off solutions. The decision-makers can sample a suitable solution based on their requirements. The experimental results show that the obtained accuracy-best neural networks can outperform hand-crafted architectures significantly, while the low-complexity and trade-off solutions can also achieve acceptable results. After the examination of these automatically generated architectures, some design principles for constructing a neural network for SOH estimation are identified for future estimating attempts.