Predictive Maintenance System for Industrial IoT: A Hybrid Deep Learning and Edge-Computing Framework

Abstract: Unplanned machinery failures in industrial environments cost global manufacturing an estimated $50 billion annually, rendering predictive maintenance (PdM) one of the most economically critical applications of Industrial Internet of Things (IIoT). Traditional threshold-based and reactive maintenance strategies fail to capture the complex, non-linear fault progression patterns exhibited by rotating machinery, compressors, and conveyor systems operating under variable load conditions. This paper proposes the Hybrid Edge-Cloud Predictive Maintenance (HECPM) framework, which integrates a Temporal Convolutional Network-Long Short-Term Memory (TCN-LSTM) ensemble for multivariate sensor time-series modelling, a Variational Autoencoder (VAE) for unsupervised anomaly detection under data-scarce conditions, and a Federated Learning (FL) orchestration layer that preserves proprietary operational data within factory edge nodes. An Explainability Module based on SHAP and attention heatmaps translates neural predictions into maintenance work-orders interpretable by floor engineers. Experiments on three publicly available benchmarks—NASA CMAPSS Turbofan, Case Western Reserve University (CWRU) Bearing, and PRONOSTIA Bearing datasets—demonstrate that HECPM achieves a Remaining Useful Life (RUL) prediction RMSE of 11.34 cycles (CMAPSS FD001), fault classification accuracy of 99.12% (CWRU), and anomaly detection F1-score of 0.963 (PRONOSTIA), outperforming all evaluated baselines. The federated deployment reduces raw sensor data transmission by 87.3% while sustaining model performance within 1.1% of centralized training, validating the framework’s industrial deployability under bandwidth and data-privacy constraints.

Keywords: Predictive Maintenance; Industrial IoT; Temporal Convolutional Network; Federated Learning; Remaining Useful Life; Anomaly Detection; Edge Computing