Abstract
Data-driven methods have become increasingly important in the ongoing advancement of condition-monitoring systems for electrical rotating machines. These methods offer streamlined implementation of effective fault detection and diagnostics. However, despite their numerous benefits, the practical application of data-driven approaches encounters obstacles such as data imbalance. Specifically, the scarcity of sufficient, reliably labeled fault data from real-world machines often hinders the development of accurate supervised learning-based condition monitoring systems.
This research investigates anomaly detection in a wound-rotor induction generator using voltage and current signature analysis of both stator and rotor windings to address the issue of imbalanced data due to either scarce or unavailable labeled fault data. The research investigated resampling methods and establishing an enhanced supervised learning approach for scarce labeled data, and suitable anomaly detection methods for fault detection in cases where labeled fault data is not available. The study employs a supervised learning-based condition monitoring system, incorporating resampling and unsupervised learning techniques. Frequency-based features, correlated with specific anomaly modes, are extracted during the signal processing phase. Finite element modeling of the wound rotor induction generator, simulating normal and various fault conditions, is performed based on the preliminary design to develop and test the anomaly detection systems. Finally, an experimental setup validates the computational findings.
The fault classification using Decision Tree, Random Forest, and k-nearest neighbors algorithms, combined with resampling techniques like SMOTE (Synthetic Minority Oversampling Technique), Tomek Links, and a combination of both is investigated. This investigation utilizes both simulated and experimental imbalanced data, focusing on supervised learning and resampling methods. A comparative analysis across different imbalance scenarios assesses the suitability of each technique for condition monitoring of a wound-rotor induction generator. Performance is evaluated using precision, recall, and F1-score metrics. Results demonstrate that the combined SMOTE and Tomek Link approach yields the best performance across all classifiers. Specifically, k-nearest neighbors, coupled with this combined resampling technique, achieved the most accurate classifications. This research is valuable for researchers and practitioners in the field of electrical machine condition monitoring. Finding the best methods for handling imbalanced fault data—a critical component of condition monitoring—will be made
Data-Driven Condition Monitoring on a Wound-Rotor Induction Generator
iii
easier with the help of the findings and comparative analysis provided here, the fault data is likely to be sparse for machines in industrial settings, particularly for those in specialized applications where condition monitoring is more likely to be necessary.
Also, the study focuses on using unsupervised learning for anomaly detection in a wound rotor induction generator. We employed stator current, voltage, and rotor current data to create anomaly detection models using isolation forest and local outlier factor methods. These models were evaluated on various simulated and real-world datasets with varying levels of imbalance. Overall, the isolation forest generally outperformed the local outlier factor for most classes, though there were some instances where the local outlier factor performed better. In some cases, the performance was similar between the two methods, but isolation forest tended to be more effective. When the data contamination rate was reduced contamination level both isolation forest and local outlier factor improved their performance across all classes, which offers 80% accuracy and recall, 90% precision, 82% F1-score, and 88% ROC AUC for isolation forest based on practical data. While for simulated data, the isolation forest presented 93% accuracy and recall, 95% precision, 94% F1-score, and 96% ROC AUC. The local outlier factor provides 76%, precision of 86%, F1-score of 77%, and ROC AUC of 82% based on simulation data, and the LOF presents 80% recall and accuracy, 90% precision, 82% F1-score, and 88% ROC AUC based on experimental data. This indicates that both techniques are effective for working with imbalanced data. We also conducted anomaly detection using a deep learning long short-term memory (LSTM) model on stator current, voltage, and rotor current data. Both simulation and experimental results showed that the LSTM model could accurately identify anomalies based on mean absolute error with the constancy accuracies. In simulations across all classes, the LSTM model's training accuracy ranged from 74.59% to 82.05%, and its validation accuracy ranged from 79.34% to 90.91%. Experimentally, the training accuracy varied between 76.74% and 85%, while the validation accuracy ranged from 77.78% to 85.71%. The investigation presents a way for improving fault detection root cause analysis that will help diagnosis with LSTM.