Intelligent Fault Diagnosis for Microgrid IoT Systems Using ML

Thesis Abstract

Smart microgrid operation increasingly relies on real-time fault diagnosis to maintain reliability, safety, and efficiency in distributed generation and IoT-enabled monitoring environments. This study addresses the challenge of timely and accurate fault detection and isolation in heterogeneous microgrid systems that integrate solar, wind, storage, and conventional generators, under varying load conditions and communication constraints. The aim is to develop an intelligent fault diagnosis framework that leverages machine learning (ML) integrated with Internet of Things (IoT) data streams to identify fault types, estimate fault severity, and localize faulty components with high accuracy and interpretability. Specific objectives include (1) constructing a comprehensive IoT-enabled microgrid dataset combining voltage, current, frequency, temperature, and actuator signals under normal and fault conditions; (2) designing a multi-stage ML pipeline that fuses temporal and spectral features from high-frequency sensors, including short-time Fourier transform and wavelet transform representations; (3) evaluating supervised, semi-supervised, and unsupervised models for fault classification and anomaly detection, with emphasis on explainability using SHAP values and feature importance analyses; (4) developing a fault localization module that integrates graph-based signal processing and causal discovery to identify faulty components within the microgrid topology; and (5) validating the framework through hardware-in-the-loop simulations and a 50-node laboratory microgrid testbed under representative contingency scenarios. Methodologically, the research adopts a mixed-methods design anchored in a systems engineering perspective and guided by the resilience theory and the cognitive reliability model to ensure robustness of diagnosis under uncertain sensor data and communication delays. The population comprises a multi-source microgrid dataset generated from a 50 kW laboratory microgrid with photovoltaic arrays, a 100 kWh battery, a diesel/gas generator set, and power electronic interfaces. A stratified sampling approach yields 20 fault scenarios per subsystem (inverter, transformer, switchgear, and energy management system) plus 200 hours of normal-operation data, culminating in a labeled dataset of approximately 1.5 million data points. Data collection instruments include high-rate sensors (2 kHz sampling) for voltage/current, PMUs, temperature probes, IoT gateways, and cyber-physical testbed controllers, complemented by archival records from a simulated utility-grade microgrid. The data analysis workflow employs feature extraction (time-domain, spectral-domain, and wavelet features), data fusion, and a hierarchical classification framework combining convolutional neural networks (CNNs) for feature learning with gradient boosting trees for interpretable decision rules. Model evaluation uses stratified k-fold cross-validation, precision-recall analysis, and receiver operating characteristic (ROC) metrics, with calibration checks to ensure reliability under data sparsity. The study incorporates transfer learning to generalize findings to diverse microgrid topologies and employs domain adaptation techniques to mitigate distributional shifts. Model interpretability is enhanced through SHAP analysis and rule-based post-processing to translate ML outputs into actionable maintenance recommendations. Expected findings include (i) a higher fault-detection accuracy (>95%) and faster detection latency (<200 ms) for transient and persistent faults compared to baseline diagnostic methods; (ii) robust fault localization with localization accuracy exceeding 90% across subsystem types; (iii) demonstrable resilience to missing data and communication delays with imputation-free performance under moderate packet loss; (iv) clear, explainable decision pathways enabling operators to understand feature contributions and rationale for fault hypotheses. The study contributes to knowledge by integrating ML-driven diagnostic algorithms with IoT-enabled microgrid observability, offering a scalable blueprint for fault management in future energy systems, and by advancing theory at the intersection of reliability engineering and data-driven cyber-physical systems. It also provides practical guidelines for deploying explainable ML in energy automation contexts and informs policy on reliability standards for ICT-enabled distributed generation networks. The main conclusion anticipates that an integrated ML-IoT fault-diagnosis framework can significantly enhance reliability and operational efficiency of microgrids, while recommendations emphasize standardized data protocols, cybersecurity hardening, real-time edge computing, and continuous model maintenance through ongoing data collection and periodic re-training to adapt to evolving grid configurations.