Sub-Cycle Sliding Window Technique for Real-Time Intelligent Electrical Fault Diagnosis at CW&PRS

Technical Summary: Sub-Cycle Sliding Window Technique for Real-Time Intelligent Electrical Fault Diagnosis at CW&PRS

Problem Statement
The electrical grid at the Central Water and Power Research Station (CW&PRS), Pune, India, faces significant operational challenges due to rapid network development and the integration of renewable energy sources. The existing protection infrastructure relies on reactive, electro-mechanical schemes that require a full Alternating Current (AC) cycle (approximately 20 ms at 50 Hz) to calculate Root Mean Square (RMS) values before isolating a fault. While safe, this latency is insufficient for modern, dynamic power infrastructures, particularly for protecting sensitive assets like large motors and Variable Frequency Drives (VFDs) from irreparable damage during transient faults. Furthermore, traditional Supervisory Control and Data Acquisition (SCADA) systems suffer from asynchronous data updates, low resolution, and high noise levels, while the inherent "class imbalance" in power system data (where normal operation vastly outnumbers fault events) complicates diagnostic modeling.

Methodology
The study proposes an intelligent, sub-cycle fault diagnosis framework designed to operate on raw, high-resolution telemetry data from Phasor Measurement Units (PMUs) and Digital Fault Recorders (DFRs) at the ESAC Centre. The methodology follows a multi-stage analytical pipeline:

1. Data Acquisition and Preprocessing: Raw data (three-phase voltage and current, frequency, and phase angles) was extracted from the ESAC Centre. An Extract, Transform, and Load (ETL) pipeline was implemented using Python (NumPy, Pandas, Scikit-Learn) to synchronize asynchronous signals, normalize magnitudes, and handle missing values.
2. Exploratory Data Analysis (EDA): The study analyzed the transition from steady-state to fault conditions, identifying distinct signatures such as current spikes (e.g., rising from ±7,500 A to >±30,000 A) and voltage sags. Statistical features (mean, standard deviation, min/max) were calculated to characterize fault classes.
3. Feature Engineering and Correlation: A correlation analysis was performed to identify dependencies and redundant variables. The analysis revealed that linear correlations between raw magnitudes and fault labels were negligible, indicating that fault detection relies on complex, non-linear patterns rather than simple thresholds.
4. Sliding Window Architecture: To bypass the 20 ms full-cycle requirement, the authors implemented a "Sub-Cycle Sliding Window" technique. Instead of waiting for a full cycle to compute RMS, the system processes a continuous stream of data using a 10-sample window (representing approximately 3.125 ms). This transforms the problem from phasor estimation to "shape recognition," allowing the model to detect deviations from the normal sinusoidal trajectory immediately upon fault inception.
5. Model Training and Classification: Various Machine Learning classifiers were evaluated, including Logistic Regression, Decision Trees, Support Vector Machines (SVM), Artificial Neural Networks (ANN), XGBoost, and Random Forest. The models were trained to distinguish between normal operation and various fault types (LG, LL, LLL) in noisy environments.
6. Post-Processing (Safety Filter): To mitigate "nervous" responses to stochastic noise, a Persistence Filter with a moving average was applied. This logic smooths transient probability spikes, ensuring alarms are only raised when a fault pattern persists for approximately 1.5 ms, balancing sensitivity with noise immunity.

Key Results
The comparative analysis of classifiers demonstrated that non-linear ensemble methods significantly outperformed linear models in this noisy, imbalanced dataset:

• Random Forest: Achieved the optimal trade-off with 98.91% accuracy, a fault recall of 95.81%, and zero false alarms. It correctly identified 2,080 fault instances.
• XGBoost: Achieved 100% accuracy on all metrics in the initial testing, though the authors note a potential risk of overfitting due to the use of synthetic data in training samples.
• Linear Models: Logistic Regression performed poorly with only 76.04% accuracy and a high rate of false alarms (1,396), confirming that linear classifiers cannot capture the non-linear decision boundaries of power system faults.
• ANN: While precise (zero false positives), it showed lower sensitivity (86.05% recall), missing 316 fault instances.

Significance and Claims
The paper claims that the proposed Sub-Cycle Sliding Window Architecture represents a transformative advancement over standard impedance-based protection. By reducing the detection latency from the conventional 20 ms to 3.125 ms, the system achieves an 84% reduction in response time. This "digital reflex" capability allows the grid to isolate transients before they cause permanent thermal damage to critical infrastructure.

The authors emphasize that their approach is specifically tailored for real-world, noisy utility environments where high-frequency, noise-free PMU data is not always available. By leveraging ensemble learning (specifically Random Forest) and a shape-recognition approach via sliding windows, the system provides a robust, low-latency early warning mechanism that maintains operational continuity while ensuring the safety of water resource assets at CW&PRS. The study concludes that the ability to handle non-linear data patterns is the decisive criterion for effective fault diagnosis in modern, stochastic grid conditions.