The application of reactive and preventive maintenance strategies to avert transformer failures and safeguard their operations have shown significant limitations in terms of high operational downtimes, over- and under-maintenance issues, maintenance fatigue and revenue loss.
The recent advancements in machine learning and artificial intelligence have positively altered the machine and equipment maintenance landscape. In this regard, predictive maintenance (PdM), in contrast to the aforementioned maintenance approaches, has laid the foundation for efficient transformer maintenance by accurately identifying incipient failures, localizing the root cause of failures and predicting the remaining useful life (RUL) thereby providing a holistic solution to the existing challenges.
This work proposed a novel Dynamic Multi-Scale Attention Convolutional Neural Network-Long Short-Term Memory (DMSA CNN-LSTM) model architecture and utilized multi-modal data fusion to address shortcomings in anomaly detection, root cause localization and RUL prediction. The significance of this work lies in providing a scalable data driven architecture that dynamically selects the most vital features across different scales (short-term and long-term) to captured intricate details especially in situations where the feature importance varies with respect to time.
This is suitable for real time deployment and aligns well with the big data landscape in providing predictive solutions for transformers at a higher performance and efficiency. Therefore, this approach leveraged on the combination of deep neural networks to provide a comprehensive diagnostic and prognostic approach to mitigate transformer faults and breakdowns thereby bolstering asset management, revenue and reliable power distribution to customers.
