2026 Theses Doctoral

Multimodal Embedded Sensing and Machine Learning-Enabled Early Warning of Lithium-Ion Battery Anomalies

Nieduzak, Tymon

Lithium-ion batteries (LIBs) have become the dominant energy storage technology for electric vehicles (EVs) and grid-scale applications due to their high energy density, long cycle life, and favorable performance characteristics. However, their widespread adoption has been accompanied by increasing safety concerns, particularly associated with thermal runaway (TR), a self-sustaining exothermic process that can lead to catastrophic failure. Existing battery management systems (BMSs) rely primarily on threshold-based monitoring of temperature, voltage, and current, which are insufficient for detecting early-stage anomalies due to delayed observability and variability across failure modes.

This dissertation presents a comprehensive and deployable framework for early detection of LIB anomalies through multimodal embedded sensing and machine learning (ML) analysis. The framework is designed as a decentralized enclosure-integrated safety monitoring layer intended to operate in parallel with conventional BMSs. Temperature microsensors are embedded within a composite battery enclosure to improve manufacturability and reduce reliance on cell level sensors. Experimental studies and simulations are conducted to characterize heat transfer behavior in composite enclosures, and a ML-based temperature inference model is developed to compensate for sensing delays associated with embedded temperature measurements.

Next, gas measurements (H₂ and CO) are incorporated as complementary sensing modalities as they can provide earlier indicators of battery degradation for certain fault mechanisms. A transformer-based autoencoder architecture is developed for sensor fusion and unsupervised LIB anomaly detection. The model is trained exclusively on multimodal normal operating data and identifies deviations through reconstruction error, enabling detection of abnormal battery behavior independent of absolute temperature thresholds. The architecture is designed to operate in real-time with an inference latency of 0.057 seconds and a computational load compatible with processing platforms commonly used in modern BMS implementations.

Experimental evaluation is performed using both normal cycling data and diverse damage induced TR datasets collected in collaboration with industry partners and procured from databases. The proposed model achieves an anomaly detection rate of 99.7% and provides early warning with a maximum lead time of 24.3 minutes compared to conventional temperature threshold exceedance. Furthermore, the model demonstrates graded severity awareness, distinguishing between low, moderate, and severe anomaly conditions. The results establish a scalable and robust approach for battery safety monitoring, combining embedded sensing, multimodal data fusion, and transformer-based intelligence to enable early and reliable detection of LIB anomalies.