Diagnóstico de falhas de estator em motores de indução trifásicos com base em trajetórias v-i derivadas da teoria de potência conservativa

Abstract

Three-phase induction motors (TIMs) are widely used in industrial environments due to their robustness, reliability, and adaptability under various operating conditions. Nonetheless, despite these characteristics, TIMs are susceptible to failures that can disrupt production lines, lead to undesired operational costs and, in more severe situations, require motor replacement. Among the possible faults, inter-turn short circuits (ITSC) in the stator windings stand out for their complexity and severity, and are particularly challenging to detect at early stages. This difficulty arises because the signals associated with incipient faults are very similar to those of normal operation, in addition to the typically rapid progression of the fault. In this context, the development of robust and efficient diagnostic systems becomes essential, since early detection can prevent progressive damage and preserve the integrity of the motor. Artificial intelligence (AI) techniques play a relevant role in the development of such diagnostic systems, especially through tools capable of classifying motor operating conditions. Among these techniques, convolutional neural networks (CNNs) are particularly noteworthy. Belonging to the deep learning domain, CNNs are primarily trained using images and are known for their pattern recognition capabilities. This study proposes the use of voltage-current (v-i) trajectories to diagnose stator faults in TIMs driven either by frequency inverters or directly from the power grid, employing CNNs. These trajectories were generated using two distinct methodologies: (i) directly from raw signals obtained experimentally in the laboratory, and (ii) through the Conservative Power Theory (CPT), which enables the decomposition of the electrical current into active (𝑖𝑎), reactive (𝑖𝑟), unbalanced (𝑖𝑢) and void (𝑖𝑣) components, each one associated with different physical phenomena in the electrical system. Consequently, the v-i trajectories of these current components also provide physical interpretability for each of the tests to which the laboratory TIM was subjected. The experimental tests covered variations in load, voltage imbalance, different operating frequencies when a frequency inverter was used, and several severities of short circuits in the stator winding. The generated trajectories were used as input to two CNN architectures: a network specifically designed for this study and a pretrained model based on DenseNet-169. Based on the evaluation metrics, it was possible to identify the best-performing trajectories for each network and each supply condition. In general, under inverter-fed operation, the v–i trajectory associated with the unbalance current achieved an accuracy of 100% in both CNN architectures, whereas the non-active component reached the same performance with DenseNet-169. Under grid supply, the traditional trajectory achieved an accuracy of 99.76% with the specific CNN. These results indicate that the proposed approach is well suited to fault diagnosis, showing that the use of v-i trajectories obtained from CPT-based current decomposition effectively contributes to the identification of stator faults in three-phase induction motors