Transformer Fault Detection

This model is created using Aargus-DIY Visual Inspection Tool. A tabular machine-learning model for early detection of electrical power transformer faults using SCADA sensor data (voltage, current, power, power factor, oil/winding temperature, and alarm indicators). The model flags an at-risk condition whenever the Oil Temperature Alarm (OTI_A), Oil Temperature Trip (OTI_T), or Magnetic Oil Gauge Alarm (MOG_A) is active, enabling predictive maintenance instead of reactive repair.

Overview

This model performs automated health monitoring of power transformers, replacing manual alarm-log review with a fast, consistent classifier trained on merged sensor telemetry.

Target: Binary fault flag — 1 if any of OTI_A, OTI_T, or MOG_A is triggered, else 0

Methodology

Data Ingestion — Five SCADA sources (CurrentVoltage.csv, Overview.csv, Power.csv, PowerFactor.csv, TotalPower.csv) merged on DeviceTimeStamp into a single time-indexed table.
Exploratory Data Analysis — Missing-value checks, descriptive statistics, correlation heatmaps, and long-term/weekly trend plots for key health indicators (OTI, WTI, ATI, OLI).
Feature Engineering — Time-based features (hour, dayofweek, month, is_weekend) added; alarm columns excluded from the feature set to avoid target leakage.
Model Training — Three classifiers trained and compared: Random Forest, XGBoost, and Logistic Regression (with feature scaling), using an 80/20 stratified train-test split.
Validation — Precision, recall, F1-score, ROC-AUC, and confusion matrix on the held-out test set.

Performance

Test Set Results (5,859 samples, 408 positive / fault instances)

Model	Precision	Recall	F1-Score	ROC-AUC
XGBoost (best)	0.998	0.993	0.995	0.996
Random Forest	0.995	0.990	0.993	0.995
Logistic Regression	0.831	0.973	0.896	0.979

Best Model: XGBoost Classifier (selected by highest F1-score)

Note: Logistic Regression trades precision for higher recall — useful if the priority is catching every possible fault at the cost of more false alarms. Tree-based models (XGBoost, Random Forest) give the best overall balance.

Usage

Installation

pip install xgboost scikit-learn pandas joblib

Load the model and predict

import joblib
import pandas as pd

# Load the trained model
model = joblib.load("transformer_xgb_model.pkl")

# feature_columns should match the training feature set
# (sensor readings + engineered time features: hour, dayofweek, month, is_weekend)
new_data = pd.read_csv("your_sensor_readings.csv")

predictions = model.predict(new_data)
probabilities = model.predict_proba(new_data)[:, 1]

print(predictions)      # 0 = normal, 1 = fault risk
print(probabilities)    # fault probability

For the Logistic Regression variant, also load and apply transformer_lr_scaler.pkl (StandardScaler) to the features before calling .predict().

Input Features

Voltage (VL1–VL31), current (IL1–IL3, INUT), power (WL1–WL3, KW, KVA, KVAR, KWH), power factor (PFL1–PFL3), oil/winding/ambient temperature (OTI, WTI, ATI), oil level (OLI), plus engineered time features (hour, dayofweek, month, is_weekend).

Use Case

Predictive maintenance for power transformers — enables early, automated detection of overheating and oil-level faults from live SCADA telemetry, reducing unplanned downtime and manual monitoring effort in industrial/utility settings.

Built With

Aargus DIY Visual Inspection Tool
scikit-learn (Random Forest, Logistic Regression)
XGBoost
pandas / NumPy
Matplotlib / Seaborn (EDA)

License

This project is licensed under the Apache 2.0 License.

Downloads last month: -; Downloads are not tracked for this model. How to track