Physics-Informed Sensor-Lean Fault Diagnosis for Lithium-Ion Batteries

Battery modules need accurate fault diagnosis with minimal sensing hardware. This paper proposes a sensor-lean framework for fault detection, localization, and classification in parallel-connected lithium iron phosphate (LFP) battery modules. The method uses a ladder network with offlinemeasured busbar parasitics to estimate cell-level electrical behavior from terminal measurements. These model-estimated features then feed a machine-learning classifier. Three early-fault surrogates are generated using physics-based simulation in liionpack/PyBaMM: connection resistance increase, anode and cathode diffusivity reduction. Under leave-one-out cross-validation (LOO-CV) on a 60-sample dataset, the classifier achieves $96.7 \%$ detection accuracy and $86.7 \%$ localization accuracy. Cross-fault validation shows the classifier can generalize across fault families with observable signatures. Hardware validation on a lab-built 1S3P module confirms successful detection and localization of a connection fault, with a voltage fit RMSE of 2.72 mV. Sensitivity analysis indicates robustness to busbar inductance calibration error, but requires busbar resistance accuracy within about $\pm 20 \%$.