State of Health Estimation using Convolutional and Bidirectional LSTM Neural Networks tuned by Bayesian Optimization

Effective battery management systems (BMS) rely on accurate State of Health (SOH) estimation to ensure safety and longevity. As battery degradation is a complex, non-linear process, data-driven approaches have become increasingly popular. This paper introduces a hybrid deep learning framework that leverages the spatial feature extraction capabilities of Convolutional Neural Networks (CNNs) and the temporal dependency modeling of Bidirectional Long Short-Term Memory (BiLSTM) networks. To maximize performance, the authors integrate Bayesian Optimization (BO) to automate the selection of network hyperparameters, such as layer counts, unit sizes, and learning rates.

The study evaluates three main architectural categories: standalone recurrent models, CNN-RNN combinations, and CNN-RNN hybrids augmented with intermediate Fully Connected (FC) layers. The authors perform an exhaustive feature engineering analysis using four key parameters—capacity (Q), voltage (V), Incremental Capacity Analysis (ICA), and Differential Voltage Analysis (DVA)—to determine the optimal input representation for each dataset. The framework was validated using three distinct datasets: two publicly available (Oxford and NASA) and one private dataset (PoliMi) developed by the authors. The models were compared using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and computational complexity measured in Floating-point Operations (FLOPs).

The results demonstrate that the CNN-FC-BiLSTM architecture consistently achieves the highest predictive accuracy for the PoliMi and Oxford datasets. For the NASA dataset, which exhibits higher noise and more aggressive degradation patterns, the CNN-BiGRU model proved more effective. The study highlights that while these hybrid models offer superior accuracy, they also demand significant computational resources. To address this, the authors identify more compact model configurations that maintain acceptable accuracy, suggesting a viable path for real-time implementation on microcontrollers.