Asian Basic and Applied Research Journal

This article introduces a previously unexplored data control approach to diagnosing inductive stove problems. Induction stoves have been used in foundries for over a century to heat and melt metal. This allows for high melting and heating speeds with optimal efficiency. However, unplanned shutdowns and errors can interfere with production and pose security risks. The proposed architecture of deep neural networks continuously monitors supply-side electrical parameters to identify electrical errors in real-time. To collect sensory and experimental data, Foundry uses a variety of devices for its energy analysers. The data samples are then marked using half-surveillance learning technology known as the local outlier factor to distinguish between normal and defective authorities. The marked data is used to train deep neural networks. The performance of the developed model is evaluated using several metrics in several advanced techniques. The results show that the deep neural network model exceeds the other classifiers and achieves an average F-measure of 0.9187. Considering the fact that neural networks act as black boxes, predictions are interpreted by Shapley Additive's explanation and locally interpretable models-logical explanations. Interpretability analysis shows that odd voltage/electric harmonic anomalies in orders 3, 11, 13, and 17 are strongly related to the identified errors, highlighting the important role of these parameters in model prediction.