Author Details
( * ) denotes Corresponding author
The Industrial Internet of Things (IIoT) has transformed modern manufacturing by generating massive volumes of multivariate sensor data, yet traditional reactive and preventive maintenance strategies continue to cause significant unplanned downtime and high operational costs. This thesis addresses these challenges by developing three novel deep learning-based predictive maintenance (PdM) models specifically designed for resource-constrained IIoT environments. The study proposes: (1) a Wasserstein GAN with LSTM discriminator (GAN-PdM) for synthetic failure data generation to overcome class imbalance, (2) a Variational Autoencoder (AE-PdM) for unsupervised anomaly detection using reconstruction error and Mahalanobis distance, and (3) a hybrid GAN-AE model (Hybrid-PdM) integrated with two optimized feature selection techniques (ISSO-FS and ERT-FS) for high-dimensional sensor streams. The models were rigorously evaluated on three benchmark IIoT datasets (AI4I 2020, NASA C-MAPSS FD001, and augmented bearing fault data) and deployed on edge devices (Raspberry Pi 5 and Jetson Nano). Experimental results demonstrate that the Hybrid-PdM model achieves state-of-the-art performance with 99.14% average accuracy, 0.51% false alarm rate (FAR), RUL prediction MAE of 3.55 cycles, and real-time inference latency of 12.47 ms while maintaining a model size under 5 MB. These outcomes significantly outperform traditional machine learning baselines and recent literature models, confirming the effectiveness of the proposed generative-reconstruction hybrid approach. The research contributes a complete, edge-ready PdM framework that reduces downtime by up to 50% and maintenance costs by up to 40%. It provides both theoretical advancements in generative and variational deep learning for industrial applications and immediately deployable tools for smart factories. Future extensions include physics-informed networks, federated learning, and digital twin integration.
Keywords
Predictive maintenance; Industrial Internet of Things (IIoT); Generative Adversarial Networks (GAN); Variational Autoencoder (VAE); Feature selection; Edge computing; Remaining Useful Life (RUL)
Abstract Views: 1
PDF Views: 5