Optimization of battery AI models for edge deployment: A systematic study of model compression techniques and hardware-aware efficiency

Abstract

Battery management systems (BMS) increasingly employ data-driven models to estimate battery state of health (SoH), but deploying such models on automotive microcontrollers remains challenging because of limited memory and computational resources. This thesis evaluates a hardware-aware workflow for transferring, compressing, and preparing a long short-term memory (LSTM)-based SoH estimation model for embedded execution. The baseline model and its original training were provided by Cognivity AI; model development is therefore outside the scope of this thesis. The study uses the provided model and its established preprocessing pipeline, with battery-level data separation for evaluation. The workflow covers PyTorch to MATLAB model transfer, numerical equivalence verification, pruning and quantization evaluation, Simulink integration, and embedded C++ code generation. Two complementary compression directions are evaluated. Structured pruning reduces the number of learnable parameters while retaining useful predictive accuracy in the reported experiment. INT8 post-training quantization (PTQ) and quantization-aware training (QAT) reduce the theoretical weight storage to approximately one quarter of the FP32 value; QAT performs better than PTQ for the unpruned baseline in the reported comparison. Compression configurations are selected with validation data and evaluated on the held-out test data only after they are frozen. The selected models are exported to Simulink for model-in-the-loop functional checks, and statically allocated C++ code is generated using Embedded Coder. The generated artifacts were cross-compiled for the Infineon KIT_A3G_TC4D7_LITE evaluation platform. These results demonstrate conversion and compilation feasibility, but do not establish INT8 arithmetic in the generated implementation, on-target execution time, power consumption, or completed hardware deployment. Those measurements remain future work.