⚡ electrical engineering

Early Anomaly-Onset Detection based on Wigner--Ville Distribution Slice Spectra: A Transmission-Grid Test Case

This study demonstrates that using full-vector Wigner–Ville Distribution Slice (WVDS) spectra for sequential anomaly detection in high-voltage grid waveforms significantly reduces false alarms compared to FFT-based and autoencoder methods, making it a superior approach for online monitoring where minimizing pre-onset false positives is critical.

Original authors: Eduardo Jr Piedad, Eduardo Prieto-Araujo, Oriol Gomis-Bellmunt

Published 2026-06-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Eduardo Jr Piedad, Eduardo Prieto-Araujo, Oriol Gomis-Bellmunt

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the electrical grid as a massive, humming orchestra. Under normal conditions, the instruments (voltage waves) play a steady, predictable tune. But when a problem starts—like a short circuit or a storm hitting a line—the music suddenly gets weird. It might develop a strange new note, a sudden wobble, or a hidden rhythm that wasn't there before.

The goal of this paper is to build a "smart listener" that can hear these weird sounds the very moment they start, before the whole orchestra crashes.

Here is how the researchers approached this, using simple analogies:

The Problem: Listening to a Moving Target

Most security systems wait until a song is finished to check if it was good or bad. But in a power grid, you can't wait. You have to listen to the music as it happens, second by second.

The Challenge: You are listening to a 128-note snippet of the song every few milliseconds. You need to decide instantly: "Is this normal, or is something wrong?"
The Trap: If your listener is too sensitive, it will scream "ALARM!" every time a violinist sneezes (a false alarm). If it's too slow, the damage is already done.

The Two Listeners: The "FFT" vs. The "WVDS"

The researchers tested two different ways to analyze these sound snippets.

1. The FFT Listener (The Standard Spectrograph)
Think of the Fast Fourier Transform (FFT) as a standard audio equalizer. It breaks the sound down into a list of frequencies: "Here is the bass, here is the treble." It's very good at spotting loud, obvious new notes.

Performance: It catches almost every problem (99% detection).
The Flaw: It gets jumpy. Because it's so sensitive, it often mistakes normal background noise for a disaster, leading to many false alarms (about 39% of the time in their test).

2. The WVDS Listener (The "Interaction" Detective)
The Wigner–Ville Distribution Slice (WVDS) is the star of this paper. Instead of just listing frequencies, it looks at how the different parts of the sound interact with each other.

The Analogy: Imagine two people talking. The FFT listener hears "Person A is talking" and "Person B is talking." The WVDS listener hears the echo between them. It notices that when Person A speaks, Person B's voice creates a specific, weird ripple in the air that only happens when they are interacting.
The Magic: In the power grid, when a fault starts, weak electrical signals interact with strong ones. The WVDS listener catches these "interaction ripples" (called cross-terms). It treats these ripples not as noise, but as a secret code that says, "Something is changing."

The Results: The "Goldilocks" Choice

The researchers tested these listeners on real data from the French power grid (RTE).

The "Fast but Jumpy" Winner (FFT): If you want to catch everything and don't mind a lot of false alarms, use the FFT. It's the fastest and most sensitive.
The "Calm and Careful" Winner (WVDS): The WVDS listener was slower to react (by a tiny fraction of a second) and missed a few more problems than the FFT. However, it was incredibly good at ignoring false alarms.
- While the FFT listener screamed "ALARM!" 39% of the time when nothing was wrong, the WVDS listener only screamed 0.69% of the time.
- Why this matters: In a real power grid, a false alarm can cause a massive, expensive shutdown of the city. The WVDS listener is like a security guard who only pulls the fire alarm when they are 99% sure there is actually a fire, rather than pulling it every time a cat walks by.

The "Teacher" and the "Students"

To make sure they were testing fairly, the researchers used a "Synthetic Teacher."

They created a computer program that learned what "normal" electricity looks like.
They used this teacher to filter out bad data and mark exactly when the "sickness" (the fault) started in the recordings.
They then taught two types of "students" (detectors) to listen:
1. Statistical Students (BND): These just calculated how far the sound was from the "normal" average.
2. Learning Students (Autoencoders): These tried to memorize the normal sound and screamed if they couldn't recreate it perfectly.

The Verdict:
The WVDS method worked best for both types of students. Even when the "Learning Students" were used, the WVDS method still reduced false alarms significantly compared to the standard FFT method.

The Bottom Line

This paper proves that by looking at how electrical waves interact with each other (using WVDS), we can build a monitoring system that is much more reliable for real-world power grids.

It trades a tiny bit of speed and sensitivity for a massive gain in trustworthiness. If you are running a power grid, you don't want a system that cries wolf every time a bird lands on a wire. You want the WVDS system: the one that stays quiet until it is absolutely sure something is wrong.

Technical Summary: Early Anomaly-Onset Detection Based on Wigner–Ville Distribution Slice Spectra

Problem Formulation
Operational disturbance monitoring in high-voltage power transmission grids requires rapid, sequential decision-making based on arriving waveform windows rather than post-event analysis of completed records. While many machine learning approaches treat this as offline classification, practical monitoring systems must detect the precise moment a signal departs from normal behavior (early anomaly-onset detection). The challenge lies in balancing detection delay, missed detections, and pre-onset false alarms, particularly when false alarms are operationally costly. Existing methods often rely on manually selected fault-frequency markers or standard spectral representations that may not fully capture transient interactions between waveform components.

Methodology
The study evaluates the Wigner–Ville Distribution Slice (WVDS) as a full-vector representation for sequential anomaly detection. Unlike traditional time-frequency analysis where cross-terms are often suppressed as artifacts, this approach leverages the bilinear midpoint interaction structure of the Wigner–Ville Distribution (WVD).

WVDS Representation: For a 128-sample sliding window, the method computes a fixed-center WVD slice. This results in a 128-dimensional vector containing the magnitude of the spectrum. Crucially, the method retains all bins, including cross-terms located at pairwise midpoint frequencies $(f_q + f_r)/2$ . The authors argue these cross-terms encode interactions between dominant and weak components, providing evidence of anomalies that standard auto-terms might miss.
Detectors: Two primary detection strategies are compared using the same input dimensions (128 vectors):
1. Baseline-Normalized Deviation (BND): A statistical score calculating the squared deviation of the current window's vector from the mean and standard deviation of normal training windows.
2. Autoencoders (AE): Deep learning models trained to reconstruct normal pre-onset windows, using reconstruction error as the anomaly score.
Comparative Baselines: The WVDS methods are benchmarked against:
- Raw time-domain windows (Raw-AE).
- Full-vector Fast Fourier Transform (FFT) magnitude spectra (FFT-BND and FFT-AE).
Evaluation Protocol: The study utilizes the French RTE (Réseau de Transport d'Électricité) database (DATA_S). A synthetic "teacher" system (an autoencoder combined with clustering) filters the dataset to ensure records begin with a normal region and annotates the onset of anomalies. The final test set comprises 436 voltage record-phase examples. Detection performance is measured by detection delay, missed detection rate, pre-onset false alarms (record-level and window-level), and online runtime per window. A strict three-window persistence rule is applied to trigger alarms.

Key Contributions

Full-Vector WVDS for Anomaly Detection: The paper introduces the use of the complete WVDS vector (without handcrafted marker selection) for online grid monitoring, preserving cross-term information as a feature rather than a nuisance.
BND Score Implementation: It proposes and evaluates a BND scoring mechanism for full-vector spectral representations, offering a non-neural-network alternative to autoencoders.
Comprehensive Trade-off Analysis: The study provides a rigorous comparison of sensitivity versus selectivity, demonstrating that WVDS offers a distinct operating point characterized by significantly lower false alarms compared to FFT-based methods, albeit with slightly reduced sensitivity.

Results
The experimental results on the filtered RTE test set highlight a clear trade-off between sensitivity and false-alarm suppression:

Sensitivity: FFT-BND achieved the highest sensitivity, detecting 99.31% of anomalies with a mean delay of 23.38 ms. WVDS-BND detected 88.07% of anomalies with a longer median delay (29.84 ms).
False Alarms: WVDS-BND significantly outperformed all other methods in suppressing pre-onset false alarms. It reduced record-level false alarms to 0.69% and window-level false alarms to 0.0061%. In contrast, FFT-BND incurred a 39.45% record-level false alarm rate.
Autoencoder Performance: The trend held for autoencoders; WVDS-AE reduced record-level false alarms (17.20%) compared to FFT-AE (23.17%) but missed more examples overall.
Runtime: All methods operated in the microsecond range. WVDS-BND was approximately 1.57 times slower than FFT-BND (10.29 µs vs. 6.54 µs per window), with the overhead primarily attributed to the construction of the WVDS representation rather than the scoring logic.

Significance and Claims
The paper claims that preserving WVD cross-term information creates a selective representation for online grid-waveform anomaly monitoring. The authors conclude that WVDS is not a universally dominant replacement for FFT features but serves a specific, critical role: it provides a conservative operating point suitable for environments where false alarms (e.g., unnecessary grid trips or alarm fatigue) are costly.

While FFT-based methods remain preferable when maximum detection sensitivity and the earliest possible alarm are the primary objectives, the WVDS approach offers a robust alternative that minimizes false positives by leveraging the physical interaction evidence encoded in the cross-terms. The study emphasizes that this selectivity is achieved without constructing a full quadratic time–frequency surface, keeping the computational cost manageable for real-time window-level inference.