Inferensys

Glossary

Feature Engineering

The process of creating lag features, rolling statistics, and gas ratios from raw sensor time-series data to improve the predictive performance of transformer diagnostic machine learning models.
Data engineer managing feature store on laptop, feature definitions visible, casual data engineering session.
PREDICTIVE DIAGNOSTICS

What is Feature Engineering?

Feature engineering is the process of transforming raw transformer sensor data into informative, predictive inputs that enhance the accuracy of machine learning models for fault detection and remaining useful life estimation.

Feature engineering is the systematic extraction and construction of predictive variables from raw time-series telemetry, such as dissolved gas concentrations and load currents, to train supervised machine learning models. In transformer diagnostics, this involves calculating gas ratios (e.g., methane/hydrogen), rolling statistics (e.g., 24-hour mean temperature), and lag features that capture temporal degradation trends invisible to instantaneous threshold monitoring.

Effective feature engineering bridges the gap between operational data historians and physics-informed models by encoding domain expertise, such as IEC 60599 diagnostic ratios or Duval Triangle coordinates, into mathematical representations. This process directly determines a model's ability to distinguish between thermal faults, partial discharge, and normal aging, making it the single highest-leverage step in building reliable predictive maintenance systems.

FEATURE ENGINEERING

Core Feature Categories for Transformer ML

The process of creating lag features, rolling statistics, and gas ratios from raw sensor time-series data to improve the predictive performance of transformer diagnostic machine learning models.

01

Temporal Lag Features

Shifting time-series data backward to capture the sequential dependency of gas generation. Lag features convert a static snapshot of DGA values into a dynamic sequence, allowing models to learn the rate of change.

  • Example: A lag-24 feature for acetylene (C₂H₂) represents the gas concentration 24 hours prior.
  • Mechanism: Enables tree-based models and LSTMs to detect trends that a single reading cannot reveal.
  • Key Insight: Rapidly rising hydrogen (H₂) with a lagged baseline is a critical indicator of partial discharge.
24-168 hrs
Typical Lag Window
02

Rolling Window Statistics

Calculating aggregate metrics over a moving temporal window to smooth sensor noise and highlight persistent shifts. Rolling statistics transform volatile raw telemetry into stable, trend-indicating features.

  • Rolling Mean: Dampens transient spikes in moisture content to reveal the true drying or wetting trend of solid insulation.
  • Rolling Standard Deviation: Quantifies the instability of a load tap changer's motor current, indicating mechanical wear.
  • Rolling Min/Max: Captures the daily peak hot-spot temperature, which directly dictates the rate of cellulose aging per IEEE C57.91.
7-30 days
Common Window Size
03

Diagnostic Gas Ratios

Engineering composite features by calculating the ratios of specific dissolved gases to classify fault types. Gas ratios are the foundational input for the Duval Triangle and IEC 60599 interpretation methods.

  • Duval Ratio 1 (CH₄/H₂): Differentiates thermal faults from electrical discharges based on the relative production of methane versus hydrogen.
  • Duval Ratio 2 (C₂H₂/C₂H₄): High values indicate high-energy arcing due to acetylene generation.
  • CO₂/CO Ratio: A critical indicator of paper degradation; a low ratio suggests pyrolysis of cellulose insulation.
3-5
Key Diagnostic Ratios
04

Differential & Delta Features

Computing the first-order difference between consecutive time steps to capture the instantaneous velocity of change. This feature type is highly sensitive to incipient faults that manifest as sudden gas spurts.

  • Delta Acetylene: A non-zero delta is an immediate red flag for active arcing, as C₂H₂ is not produced by normal aging.
  • Rate of Rise: Calculating the ppm-per-day increase in combustible gases to trigger alarms before absolute thresholds are breached.
  • Application: Essential for online DGA monitors where trending velocity is more actionable than static concentration.
< 1 ppm/day
Normal Rate of Rise
05

Fourier & Spectral Features

Transforming time-domain signals into the frequency domain to extract periodic patterns invisible in raw waveforms. Spectral features are critical for analyzing vibration and acoustic partial discharge data.

  • Fast Fourier Transform (FFT): Decomposes vibration signals from load tap changers to identify specific mechanical resonance frequencies indicating misalignment.
  • Power Spectral Density: Quantifies the energy distribution of acoustic emissions to distinguish internal partial discharge from external corona.
  • Cepstral Analysis: Used to detect periodic echoes in Frequency Response Analysis (FRA) traces, revealing winding deformation.
50/60 Hz
Fundamental Harmonic
06

Interaction & Cross-Features

Creating new features by multiplying, dividing, or combining two or more independent variables to capture non-linear physical relationships. Interaction features encode domain expertise directly into the model.

  • Load × Ambient Temperature: A combined thermal stress index that better predicts hot-spot temperature than either variable alone.
  • Moisture × Oxygen: Models the synergistic catalytic effect that drastically accelerates cellulose depolymerization.
  • Gas Ratio × Load: Contextualizes a high acetylene ratio with the transformer's operating state to suppress false positives during transient heavy loading.
SHAP
Interaction Validation
FEATURE ENGINEERING FOR TRANSFORMER DIAGNOSTICS

Frequently Asked Questions

Clear, technical answers to the most common questions about building predictive features from raw transformer sensor data for machine learning models.

Feature engineering is the systematic process of transforming raw time-series sensor data—such as dissolved gas concentrations, load current, and top-oil temperature—into informative numerical representations that improve the predictive performance of machine learning models. In transformer diagnostics, this involves creating lag features (prior hourly or daily readings), rolling statistics (moving averages and standard deviations over configurable windows), and domain-specific gas ratios (e.g., methane/hydrogen, acetylene/ethylene) that encode physical fault signatures. The goal is to expose the underlying thermodynamic and chemical degradation patterns to the algorithm, enabling it to distinguish between normal operational transients and incipient faults like partial discharge or thermal overheating. Without rigorous feature engineering, even advanced models like gradient-boosted trees or LSTMs will fail to generalize beyond the training data, as raw sensor values alone rarely capture the temporal dependencies and rate-of-change dynamics critical for early failure detection.

Prasad Kumkar

About the author

Prasad Kumkar

CEO & MD, Inference Systems

Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.

His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.