Metodología para la detección y clasificación en tiempo real de muescas de tensión

Abstract

Voltage notches, characterized by short-duration, periodic deviations in voltage waveforms often accompanied by high-frequency oscillations, represent a critical yet understudied type of Power Quality Disturbance (PQD) prevalent in industrial low-voltage networks. Caused by the normal operation of line-commutated power converters, these disturbances can trigger resonances, cause equipment malfunction, and accelerate aging, leading to significant economic losses. This thesis proposes and validates a novel methodology for real-time detection, classification, and severity assessment of voltage notches, addressing substantial gaps identified through a comprehensive literature review. A systematic review revealed that existing approaches offer limited focus on both non-oscillatory and oscillatory notches and lack robust, computationally efficient detection and classification methods. To overcome these limitations, a flexible MATLAB/Simulink-based simulation platform was developed to synthetically generate voltage and current waveforms associated with commutation notches. By systematically varying parameters such as short-circuit power, firing angle, snubber circuits, and background voltage unbalance and distortion, the platform generates realistic signals validated against field measurements, thereby providing a solid foundation for the development of the methodology. The novel methodology integrates the Space Phasor Model (SPM) and Fourier Descriptors (FDs) within a deep autoencoder framework, trained exclusively on notch-containing signals. FDs are computed from the SPM of voltage and current, and characteristic FDs, derived from pulse rectifier harmonic theory, are identified as reliable spectral signatures for notch detection. A novel anomaly detection strategy, based on autoencoder reconstruction errors (where notch signals exhibit low reconstruction error and non-notch signals yield higher error), yields over 99% accuracy on simulated data and 96% on field measurements. Subtype classification (non-oscillatory vs. oscillatory) is performed using a linear support vector machine trained in the autoencoder’s latent space, achieving over 97% accuracy. Severity assessment is performed using physically interpretable indices derived from the SPM radius reconstructed exclusively with characteristic FDs (“notch-only” SPM), allowing classification into mild, moderate, significant, or severe categories. These insights support targeted mitigation strategies. Finally, real-time performance evaluations confirm that the comprehensive detection, classification, and severity assessment pipeline consistently operates within a single power frequency cycle, even at high sampling rates, demonstrating the methodology’s suitability for embedded or edge-computing applications. Overall, this thesis significantly advances PQD analysis by delivering a robust, interpretable, and computationally efficient framework for voltage notch detection and classification.