The question
An electrical grid rarely fails all at once. It sags, swells, flickers, accumulates harmonics, and produces brief transients — each leaving a distinct signature in the voltage waveform. Those signatures overlap, arrive superimposed, and are non-stationary by nature: a sag that lasts three cycles is not a stationary phenomenon that can be characterized by a single spectrum. How should such an event be represented so that a classifier can tell it apart from the disturbance it most resembles?
Overview
This project develops a signal processing and scientific machine learning framework for representing and classifying power quality disturbance signals. It treats representation as the substantive problem rather than a preprocessing step. The classifier matters, but the decision that determines performance is which features are extracted from a non-stationary waveform in the first place.
The framework is built on disturbance classes defined in IEEE 1159-2019, which provides a standardized vocabulary and makes results comparable across studies instead of dependent on locally generated datasets. Feature extraction uses wavelet packet decomposition with energy-entropy signatures — a representation the group has developed in mechanical fault diagnosis, where the underlying problem is formally similar: identifying a transient event embedded in a non-stationary signal.
Classification is treated comparatively rather than as an argument for a single architecture. Benchmarking several classifiers under identical feature sets separates gains attributable to representation from gains attributable to model capacity — a distinction frequently collapsed in the applied literature, where a strong classifier can mask a weak feature set.
The honest framing is that PQD classification is a well-populated field with many published pipelines reporting high accuracy. The contribution here is methodological transparency: standardized classes, a feature representation with physical interpretation, and a benchmark that makes the source of performance attributable.
Approach and methods
Disturbance signals are modeled parametrically according to IEEE 1159-2019 categories, which allows controlled variation of magnitude, duration and harmonic content, and produces ground truth that is exactly known — a condition rarely available with field recordings.
Wavelet packet decomposition then splits each signal across a frequency-localized tree. Rather than using coefficients directly, the pipeline computes energy and entropy measures per node: energy indicates where in the time-frequency plane the disturbance deposits its content, and entropy indicates how ordered that deposition is. Together they compress a long non-stationary waveform into a low-dimensional vector that remains physically interpretable.
Classification is then performed with tabular architectures — FT-Transformer among them — evaluated against conventional baselines on identical features, so that comparison isolates the effect of the model rather than confounding it with preprocessing choices.
Team
- Marcus Varanis — scientific coordination; time-frequency representation and feature design.
- Rodrigo Carvalho — master’s researcher in scientific machine learning; pipeline development and classifier benchmarking.
Additional members of the Nonlinear Dynamics and Signal Processing Group contribute to the signal processing and machine learning components.
Related publications
The methodological foundation of this project rests on the group’s published work in wavelet-based analysis of non-stationary signals:
- Varanis, Pederiva. Statements on wavelet packet energy-entropy signatures and filter influence in fault diagnosis of induction motor in non-stationary operations. Journal of the Brazilian Society of Mechanical Sciences and Engineering, 2018.
- Varanis, Silva, Balthazar, Pederiva. A tutorial review on time-frequency analysis of non-stationary vibration signals with nonlinear dynamics applications. Brazilian Journal of Physics, 2021.
Current directions
Consolidating the benchmark across the full set of IEEE 1159-2019 classes, including superimposed disturbances, where single-label classification ceases to be the appropriate formulation. Work continues on how far energy-entropy features remain discriminative as disturbance classes are combined, and on how sensitive the pipeline is to sampling rate and noise floor.
For prospective students
This project spans P2 — Nonstationary Signal Processing and P3 — Scientific Machine Learning. Useful background: signals and systems, Fourier analysis, and Python. Familiarity with wavelets is helpful but is normally acquired within the project.
A student working here builds the full path from synthetic signal to classification decision: generating standardized disturbances, decomposing them, extracting features, and then testing whether a classifier’s accuracy survives conditions it was not tuned for. The instructive part is usually the failure cases — the disturbance pairs that remain confusable, and what that says about the representation rather than about the model.