MVLSC Home · Issue Contents · Forthcoming Papers
A Framework Proposal to Deal with Uncertainty in Operational Conditions Through an Iiot Monitoring System Powered by Deep Learning
Pablo Aqueveque, Pedro Pinacho-David, Felipe Monsalvez, Anibal S. Morales and Ernesto Guerra-Vallejos
Condition monitoring and predictive maintenance often struggle with the variability and uncertainty of operational conditions on industry applications, beside the high costs of collecting contextual data. This study presents an ensemble model that integrates a Long Short-Term Memory (LSTM) anomaly detector with a Convolutional Neural Network (CNN) classifier, enabling automatic labeling and learning of new conditions. In this paper, the model is tested on twelve operating scenarios of a vibratory screener under different load conditions, after an initial training with supervised learning on datasets representing two and three pre-classified operating conditions. The results show distinctiveness in the learned classes, achieving completeness and homogeneity scores of 0.93 and 0.59 in the first training, improving the results to 0.92 and 0.79, correspondingly, in the second round of training. The proposed ensemble model consolidate multiple conditions into unified categories, forming a mix of the original classes. Statistical analysis of these learned classes provides insights about equipment areas with a higher likelihood of failure. This method provides a reliable and practical solution for asset monitoring in complex industrial applications where failures involve multiple factors and their root causes are not so evident or remain unclear.
Keywords: Condition Monitoring, Predictive Maintenance, Anomaly Detection, Evolving Classification, Ensemble, Uncertainty, Deep Learning, Self-learning