Advanced Machine Learning Techniques for Time Series Analysis of Large-Scale, Real-World, Noisy Data

Abstract

The reliable operation of modern power distribution networks is essential for their safety and efficiency. Partial discharges (PDs) within the insulation of covered conductors (CCs) threaten network integrity because they can cause insulation degradation, outages, and safety incidents. This dissertation addresses the detection and classification of PDs using contactless antenna-based measurements. Such data are voluminous, highly noisy, imbalanced between PD and non-PD classes, and often contain only subtle, transient PD signatures masked by interference. Existing algorithms struggle under these conditions, resulting in low accuracy, high false-alarm rates, and difficult deployment in resource-constrained environments.

This work therefore introduces new machine learning methods that are noise-resilient and optimized for deployment on edge computing platforms. Emphasis is placed on low power consumption, high classification accuracy, and the minimization of false positives, which are critical for power system reliability.

The main contributions of this dissertation are as follows:

1. A novel stacking-ensemble deep learning architecture for raw one-dimensional time-series data. It combines one-dimensional convolutional neural networks (1D-CNNs) for robust feature extraction with autoencoders for anomaly detection. Together with a custom denoising algorithm, it delivers significantly better performance than existing methods and a wide range of contemporary classifiers when classifying PDs directly from unprocessed, noisy signals.

2. A hierarchical ensemble of two-dimensional CNNs applied to PD spectrograms, designed for low-latency, real-time classification on resource-constrained devices (such as Google Edge TPU or NVIDIA Jetson) while maintaining excellent accuracy and energy efficiency.

3. A lossy data compression algorithm based on a deep ResNet autoencoder with integrated anomaly correction and an adaptive loss, enabling efficient transmission of large-scale PD data over unreliable networks (e.g., 2G GSM) while preserving diagnostically critical features and transient anomalies that standard compressors tend to discard.

4. The creation, meticulous validation, and public release of a large-scale dataset of contactless PD measurements from real covered conductors, with both automatic and expert annotations, serving as a reference for reproducible research and benchmarking of new methods.

5. An evaluation of conditional generative adversarial networks (cGANs) for synthesizing realistic PD spectrograms, showing that cGANs effectively augment training data, mitigate class imbalance, and lead to statistically significant improvements in classifier generalization.

6. Advanced on-device preliminary screening using lightweight two-dimensional CNNs, applied to spectrograms or newly proposed two-dimensional histograms (derived from signal amplitude and steepness). These methods achieve high recall while substantially reducing transmitted data volume and energy consumption of remote monitoring units.

The proposed approaches substantially enhance the reliability and accuracy of fault diagnosis in covered conductors, including the challenging detection of high-impedance faults that often precede insulation failure, and contribute broadly to time-series analysis of noisy, rare-event real-world data. Future work will focus on further optimization, integration, and field validation, including high-confidence remote (server-side) applications and the development of adaptive, intelligent monitoring systems to strengthen the dependability of modern power systems.