Lithium-Ion Battery Diagnostics

Battery diagnosis is an inference problem. The quantities of greatest interest cannot usually be measured directly during operation: state of charge, state of health, available power, remaining useful life, and the dominant ageing mechanisms must be estimated from incomplete observations.

This topic was also the subject of my doctoral research. My PhD thesis, focused on lithium-ion battery diagnosis using electrochemical noise analysis, is available from the French open-access archive HAL: PhD thesis on HAL.

What must be estimated

State of charge describes the fraction of usable capacity available under a defined reference condition. State of health describes how the battery has changed relative to an initial or nominal condition. Neither concept has a single universal definition: capacity loss, resistance increase, power capability, and safety margins may evolve differently.

A useful diagnostic method must therefore state exactly which property it estimates, under which conditions, and with which uncertainty. A percentage without a reference procedure is not a physical result; it is decoration.

Available measurements

Operational diagnosis commonly relies on:

  • current, terminal voltage, and temperature;
  • charge and discharge history;
  • rest periods and relaxation behaviour;
  • capacity or pulse tests performed when available;
  • impedance-related indicators;
  • models linking internal states to measurable quantities.

Each measurement has limits. Voltage depends on chemistry, temperature, current, hysteresis, and relaxation. Coulomb counting accumulates sensor and initialisation errors. Resistance depends on the pulse definition and operating point. Capacity tests are informative but intrusive and time-consuming.

Ageing is not one-dimensional

Lithium-ion cells age through interacting mechanisms whose importance depends on chemistry and use. Loss of cyclable lithium, loss of active material, electrolyte degradation, interface growth, and mechanical changes can produce partly similar electrical effects.

Consequently, a diagnostic indicator may correlate with ageing in one dataset without uniquely identifying a mechanism. Extrapolating it to another cell format, temperature range, or duty cycle requires new evidence.

From signals to diagnosis

A defensible workflow separates four stages:

  1. measurement and calibration;
  2. preprocessing with documented assumptions;
  3. extraction of physically or statistically motivated indicators;
  4. validation against independent reference quantities.

Machine learning can combine many indicators and capture nonlinear relations, but it also creates new failure modes: leakage between training and test data, cell-specific shortcuts, poorly represented operating conditions, and overconfident predictions outside the training domain.

Validation

Randomly splitting individual samples is often insufficient because adjacent measurements from the same cell are strongly related. Validation should be designed around the intended use: unseen cycles, unseen cells, unseen ageing trajectories, or unseen operating conditions.

Performance must be reported together with uncertainty, reference-method quality, and failure cases. The final question is not whether an estimator has a small average error, but whether its errors are understood well enough for the decision it is intended to support.

Selected publications

  • S. Martemianov, F. Maillard, A. Thomas, P. Lagonotte, L. Madier, Noise diagnosis of commercial Li-ion batteries using high-order moments, Russian Journal of Electrochemistry, 52(12), 1122–1130, 2016. DOI
  • S. Martemianov et al., New methodology of electrochemical noise analysis and applications for commercial Li-ion batteries,Journal of Solid State Electrochemistry/, 19, 2803–2810, 2015. DOI
  • F. Maillard, S. Martemianov, A. Thomas, N. Adiutantov, P. Lagonotte, L. Madier, Measurements and Signal Processing of Li-ion Electrochemical Noise, 2015. DOI