Robust deep neural network-based internet of things for power transformer fault diagnosis under imbalanced data and uncertainties

• Using SMOTE-ENN to balance data and enhance transformer DGA fault diagnosis. • Applying RobustScaler to preprocess data, improving accuracy against outliers. • Proposing a DNN with RobustScaler, SMOTE-ENN, and an optimal optimizer. • Integrating the deep learning model with IoT for fault visualization. • IoT-based monitoring analyzes gas data for faster transformer fault detection. One of the most vital components of power systems is power transformers, which provide an essential link in the chain of other devices used to supply electricity to consumers. According to the literature, the Duval pentagon method (DPM) is one of the most accurate and reliable dissolved gas analysis (DGA) interpretation methodologies. However, implementing large amounts of data in DPM is still challenging and has several limitations. To overcome these limitations, this paper introduces a robust deep neural network (DNN) method for precise DGA monitoring. Another merit is the proposal of synthetic minority over-sampling technique-edited nearest neighbor (SMOTE-ENN) preprocessing to eliminate noise from the imbalanced dataset, resulting in cleaner merged DGA samples. Furthermore, a unique RobustScaler technique is employed to maintain high performance against uncertain data noise. To visualize transformer faults remotely and enhance the acceleration of decision-making regarding the transformer status, this paper utilizes an industrial Internet of Things (IoT) platform. Specifically, the designed deep learning model is hybridized with an IoT platform to analyze the transferred DPM dataset of the gases concentration and send the classification results using the IoT gateway to the cloud for visualizing the detected fault on the IoT dashboard. The empirical results display that the proposed method outperforms several state-of-the-art approaches. The proposed method achieves satisfaction in diagnosing faults for the assessment dataset, with an accuracy of 98.19 %. Besides, the obtained results illustrate the effectiveness of the proposed model against uncertainty noise up to 20 % with a superior prediction diagnosis of the transformer faults.