# Hybrid deep learning framework for real-time fault detection in squirrel-cage induction motors

## TL;DR

This paper introduces a hybrid deep learning framework for detecting faults in squirrel-cage induction motors in real time, using sensor data and achieving high accuracy.

## Contribution

The novel contribution is a hybrid deep learning framework combining CNN and GRU/LSTM models for real-time fault detection in SCIMs with high accuracy and efficiency.

## Key findings

- Hybrid models CNN-GRU and CNN-LSTM achieved 92.57% and 92.27% classification accuracy, outperforming other models.
- The framework efficiently captures temporal and spatial features, suitable for real-time deployment.
- The system detects various fault types using real-time sensor data like torque, speed, and currents.

## Abstract

The Fourth Industrial Revolution has heightened the demand for intelligent and reliable predictive maintenance systems in industrial environments. This study proposes a hybrid deep learning-based framework for real-time fault detection in Squirrel-Cage Induction Motors (SCIMs). Utilizing eight deep learning architectures—CNN-GRU, CNN-LSTM, LSTM, BiLSTM, Stacked LSTM, GRU, CNN, and ANN—the framework was trained and tested on a comprehensive dataset comprising one million samples, evenly divided between healthy and faulty motor conditions. Hybrid models, particularly CNN-GRU and CNN-LSTM, achieved classification accuracies of 92.57% and 92.27%, respectively, outperforming the other baseline models across precision, recall, and F1-score by effectively capturing both temporal and spatial features. Beyond classification accuracy, the hybrids further demonstrated computational efficiency in terms of inference time, latency, and throughput, validating their suitability for real-time deployment. The system analyzes real-time sensor data, including torque, speed, power, and stator/rotor currents, to identify various fault types such as short circuits, overloads, mechanical failures, and open circuits. Developed in MATLAB Simulink, the framework demonstrates high accuracy and scalability for real-time deployment. While results are promising, the claims are positioned within the scope of the evaluated models, as direct benchmarking with state-of-the-art methods was not within the present scope. The framework demands substantial computing power and annotated datasets, yet it represents a step toward intelligent, self-aware industrial systems. Future work will focus on model optimization, deployment in resource-constrained environments, and validation with real-world noisy industrial data, explicitly considering sensor drift, varying load conditions, and fault severity levels across diverse motor types and operational scenarios. In addition, since bearing faults account for a significant share of induction motor failures in practice, it will be a key priority to ensure comprehensive industrial applicability.