Battery life prediction

Predicting how many cycles a lithium-ion cell has left from its charge curves, with a model small enough for an embedded board.

Reimplemented and extended a published method; drafted a patent concept for phone-based impedance diagnostics.

Extract the incremental-capacity curve

dQ/dV is computed from voltage and charge, smoothed with a Kalman filter, and restricted to the 3.8 to 4.2 V region where degradation shows.

Seven features per cycle

Peak value and peak voltage, as in earlier work, plus the area under the curve, peak width, peak skewness and two cycle-to-cycle change features.

A three-cycle window

The features of cycles t−2, t−1 and t are stacked into one 21-number input, so the model sees a trend, not a snapshot.

Robust scaling and split

RobustScaler for resistance to outliers, then a shuffled 60/40 train and test split.

Butterfly optimisation

24 candidate solutions over 100 iterations search for the input weights of the network.

Extreme Learning Machine

32 sigmoid neurons. The output weights are solved in one step with the Moore-Penrose inverse, with no backpropagation.

Remaining useful life

Remaining cycles are derived from capacity against an end-of-life threshold of 1.4 Ah.