Predictive AI Maintenance of Distribution Oil‐Immersed Transformer via Multimodal Data Fusion: A New Dynamic Multiscale Attention CNN‐LSTM Anomaly Detection Model for Industrial Energy Management
Elvis Tamakloe, Benjamin Kommey, Jerry John Kponyo, Eric Tutu Tchao, Andrew Selasi Agbemenu, Griffith Selorm Klogo
Abstract
ABSTRACT Reactive and preventive maintenance strategies have been applied to avert transformer failures and safeguard their operations. However, these approaches have limitations of high operational downtimes, over‐ and under‐maintenance issues, maintenance fatigue and revenue loss. The advancements in machine learning and artificial intelligence have positively altered the machine and equipment maintenance landscape. Thus, predictive maintenance (PdM), in contrast to the above‐listed maintenance approaches, has laid the foundation for improving transformer maintenance by identifying incipient failures to solve the existing challenges. Recent developments in predictive maintenance of distribution power transformers have made great strides, but to solve the current challenge of accurate fault identification, this study proposed a new model architecture (DMSA CNN‐LSTM) using multimodal data fusion to address anomaly detection. A classification accuracy, F1‐score, precision and recall of 0.9917, 0.9714, 0.9781 and 0.9647, respectively, were produced on a fused multimodal dataset at a computational time of 619.898 s. The performance was afterwards evaluated against other state‐of‐the‐art benchmark models. The significance of this study lies in providing a scalable data‐driven architecture suitable for real‐time deployment in providing predictive solutions for transformers at a higher performance efficiency. This approach leverages deep neural networks that provide a comprehensive diagnostic and prognostic approach to mitigate transformer faults and breakdowns.