Identificação de falhas em motores de indução trifásicos usando rede neural transformer convolucional

Abstract

This work presents a hybrid approach using Convolutional Neural Networks (CNN) and Transformers for fault diagnosis in three-phase induction motors, focusing on the detection and classification of the severity of broken bar faults based on current and voltage signals. Electrical Signature Analysis (ESA), widely used in motor monitoring, offers several advantages. However, ESA-based techniques traditionally rely on spectral transformations, which can result in high computational cost and reduced generalization capability. CNNs can extract discriminative features directly from raw data, eliminating the need for preprocessing steps. The proposed study integrates CNNs with the attention mechanism of Transformers, which captures spatiotemporal dependencies in the data. The Convolutional Transformer Neural Network (CTNN) achieved approximately 97% accuracy when using the entire dataset, significantly outperforming classical machine learning algorithms such as Random Forest and k-Nearest Neighbors (KNN), which obtained 90% and 86% accuracy, respectively. The CNN, tested under similar conditions, achieved 96% accuracy. Compared to other methodologies involving multiple preprocessing steps and transformations to the frequency domain, the proposed approach achieves similar results, close to 100% accuracy, while being simpler, more efficient, and with greater generalization capability. Additionally, the methodology employs a reduced sampling rate, approximately six times lower than the original sampling rate of the dataset, contributing to computational cost reduction without compromising performance