Hybrid Machine Learning and Deep Learning Framework for Predictive Maintenance in Industrial IoT Systems

Abstract

In the Industrial Internet of Things (IIoT) systems, predictive maintenance is essential when it comes to minimizing downtime of equipment and the cost of operations. Conventional rule-based maintenance methods have a low degree of flexibility and a high false-positive rate, but standalone machine learning (ML) or deep learning (DL) models can be ineffective in handling higher-order patterns in industrial sensor data. In this paper, we suggest using a hybrid framework with the incorporation of the superior DL structures (CNN-LSTM) and the unsupervised anomaly detection (Isolation Forest + One-Class SVM) to provide strong predictive maintenance. The framework is tested on NASA turbofan engine degradation data, and its performance measures have better performance: 96.8% accuracy during supervised learning, 94.2% accuracy during CNN-LSTM hybrid model and 92.5% accuracy during unsupervised anomaly detection. The hybrid method proposed in the paper has 8-12% better performance than the baseline ML models and 5-7% better performance than the standalone DL models. The hybrid system proves to be better in generalizing, fewer false negatives (recall has increased to 94%), and more decision robustness due to model fusion. The study will deal with the key IIoT issues such as heterogeneity of features, temporal constraints, and class imbalance that will offer practitioners with a scalable, production-ready, industrial predictive maintenance system.