Electrochemical Noise
Electrochemical noise is the spontaneous fluctuation of electrical quantities such as potential or current around their mean behaviour. Unlike an imposed perturbation, it is observed during the natural evolution of the system. This makes it attractive for non-intrusive monitoring—and dangerously easy to overinterpret.
What is actually measured
The recorded signal contains several contributions:
- fluctuations generated by electrochemical processes;
- sensor and amplifier noise;
- quantisation and acquisition artefacts;
- electromagnetic and environmental disturbances;
- slow drift caused by temperature or state evolution;
- disturbances introduced by the experimental arrangement itself.
Before assigning physical meaning to a feature, the complete measurement chain must be characterised. The noise floor, bandwidth, sampling frequency, anti-alias filtering, grounding, shielding, and sensor stability are part of the experiment, not administrative details.
Time and frequency analysis
Simple statistics—variance, root-mean-square amplitude, distribution shape, and correlation—provide an initial description. Spectral analysis then shows how fluctuation energy is distributed with frequency, provided that the stationarity and windowing assumptions are defensible.
Electrochemical signals often evolve with operating state. A single spectrum may hide intermittent or scale-dependent behaviour. Time-frequency methods, including wavelet decompositions, can preserve localisation and separate features across scales.
Wavelets and feature extraction
A wavelet decomposition represents a signal through coefficients associated with both scale and position. Energy, variance, entropy, or distributional features can then be calculated for selected scales.
The method is powerful but not automatic. Results depend on the wavelet family, boundary treatment, decomposition depth, sampling rate, and prior detrending. A classifier that performs well after trying many alternatives may simply encode the selection process unless those choices are validated independently.
Diagnostic interpretation
My doctoral research investigated electrochemical-noise features for lithium-ion battery diagnosis. The central challenge was not producing a separation between groups, but determining whether that separation was stable, repeatable, and connected to the physical condition of the cells.
Unsupervised learning can reveal structure without predefined labels, yet a cluster is not a degradation mechanism. Its interpretation requires comparison with reference measurements, experimental controls, and knowledge of the system.
A cautious conclusion
Electrochemical noise can provide complementary information, particularly when active perturbation is undesirable. It should not be treated as a magical fingerprint. Its scientific value depends on metrology, transparent processing, repeatability, and explicit limits on what the observed features can prove.
Thesis and related publications
Electrochemical-noise analysis was the central methodology investigated during my doctoral research. The complete PhD thesis is available from the French open-access archive HAL:
Selected communications focusing on electrochemical-noise measurement, signal processing, and diagnostic methodology: