Research

A battery-management system needs more than a voltage reading. Useful diagnosis requires estimates of state of charge, state of health, available power, and the evolution of degradation under uncertain operating conditions.

These quantities are not directly measurable. They must be inferred from current, voltage, temperature, usage history, and models whose validity depends on chemistry, ageing mode, and operating range.

My doctoral work studied whether spontaneous electrical fluctuations could provide additional diagnostic information about lithium-ion cells. The aim was not to replace established techniques, but to investigate a complementary and potentially non-intrusive source of information.

Electrochemical noise refers to naturally occurring fluctuations in measured electrical quantities such as voltage or current. These signals may contain information about underlying electrochemical processes, but their low amplitude makes them particularly sensitive to instrumentation, environmental disturbances, acquisition conditions, and preprocessing choices.

A credible analysis therefore requires more than plotting a spectrum. It must address:

• the measurement chain and its own noise floor;
• stationarity and sampling assumptions;
• removal of trends without destroying relevant information;
• comparison in the time, frequency, and time-frequency domains;
• repeatability across cells and operating conditions;
• physical interpretation of extracted indicators.

Wavelet decompositions are useful in this context because they preserve a notion of scale and localisation in time. They can support feature extraction, but they do not by themselves prove that a feature is related to a particular degradation mechanism.

I also study proton-exchange membrane fuel cells, particularly gas transport, water management, and dead-end operation. In a dead-end configuration, inert species and liquid water may accumulate, so purge events become part of the system dynamics rather than a secondary control detail.

An open and useful model should connect mass balances, membrane hydration, pressure losses, electrochemical losses, and purge logic while remaining identifiable from realistic measurements. The right level of complexity depends on the question: a control model, a diagnostic model, and a detailed multiphysics model should not be confused.

The methods I use or investigate include:

• experimental data analysis and uncertainty assessment;
• time-domain, spectral, and wavelet-based signal processing;
• statistical feature extraction and unsupervised learning;
• physical and reduced-order modelling;
• numerical analysis and scientific visualisation;
• Python-based reproducible workflows.

A result is not reproducible merely because its source code is available. A complete workflow should identify the data, units, preprocessing, software environment, parameters, random seeds where relevant, and the exact artefacts used to produce each figure or table.

My preferred approach is to keep raw data immutable, express transformations as version-controlled code, test important invariants, and generate results from documented commands. Conclusions should state both what the evidence supports and what it does not.