Applied Deep Learning for Predictive Maintenance in Industrial IoT Systems

Abstrakt

Industrial IoT deployments generate continuous sensor streams from rotating machinery that, in principle, allow early detection of impending equipment failure, yet translating raw sensor data into reliable maintenance predictions remains challenging under real factory floor conditions with variable operating loads. This study applies a deep learning approach combining convolutional feature extraction with recurrent temporal modeling to predict impending bearing failures from vibration and temperature sensor streams collected across multiple industrial production lines. Rather than training on idealized failure progression data, the model is trained on sensor histories capturing normal operating variability across shift changes and load fluctuations, aiming to reduce false alarms triggered by benign operational variation rather than genuine degradation. Evaluation against maintenance logs spanning an extended production period shows the model achieving earlier failure detection than existing threshold-based alarm systems, with a meaningfully lower false alarm rate under variable load conditions. We examine model behavior across different machinery types, finding that transfer learning from data-rich machine categories improved prediction reliability for machinery types with comparatively limited historical failure data. These findings support practical deployment of deep learning predictive maintenance models in industrial settings with heterogeneous equipment fleets and uneven historical data availability.