Energy-Aware Predictive Maintenance in Industrial Edge Systems Using Pruned LSTM Networks for Sensor-Based Time-Series Data

Abstract

In modern industrial environments, predictive maintenance has become a vital strategy for ensuring operational reliability, reducing downtime, and optimizing energy utilization. However, existing deep learning (DL) approaches such as CNN, GRU, and hybrid architectures, while accurate, often suffer from high computational complexity and energy consumption, making them unsuitable for real-time edge deployment. To address these limitations, this study proposes an Energy-Aware Predictive Maintenance framework using Pruned LSTM Networks for sensor-based time-series data, designed specifically for industrial edge systems. The model employs structured pruning techniques to reduce redundant parameters and computational overhead while preserving the temporal learning capability of LSTM. The proposed approach was implemented using Python and TensorFlow on the Kaggle Industrial Equipment Monitoring Dataset, which contains multi-sensor readings representing normal and faulty machine states. Experimental results show that the Pruned LSTM model achieves a 98.8% accuracy, marking an increase of approximately 6–7% over conventional models like GRU and CNN, while reducing energy consumption by nearly 40% compared to baseline methods. This improvement demonstrates the model’s ability to maintain high precision and reliability under resource constraints. The proposed framework establishes a strong foundation for real-time edge-based predictive analytics, offering both energy efficiency and predictive robustness. In the future, the model will be extended with adaptive transfer learning and federated edge optimization to enable scalable and cross-domain industrial applications, driving the next generation of intelligent and sustainable maintenance systems.