Bearing Failure Prognostics Using Recurrent Neural Networks: A Spectral Data Based Architecture

• Renato Cosin USP
• Leo Sampaio Ferraz Ribeiro UNICAMP
• Kalinka Castelo Branco USP

Resumo

Predictive maintenance is crucial for reducing costs in the industry. With the widespread use of internet-connected sensors in industrial equipment, state-of-the-art predictive maintenance algorithms have become a prolific field for innovation. In this paper we present a new deep learning solution for predicting the remaining useful life of bearings. Bearings are widely used in industrial equipment, and their failure prognosis is highly relevant. The developed model takes spectrograms of vibration signals from bearings as input and computes their remaining use-ful life as output using a combination of Convolutional and LSTM neural networks. The model hyperparameters were optimized using the Hyperband algorithm. The dataset used originates from a large accelerated degradation experiment aimed at evolving bearing failure prognosis techniques, made publicly available as part of the IEEE PHM 2012 Data Challenge. The optimized model presented satisfactory results. In addition to reducing maintenance costs and downtime, the potential application in IIoT systems for online monitoring guided the architecture and data processing flow definition. Using a proposed criterion, the model successfully prescribed component replacement before failure in all test cases. While 20% of the maintenance was premature, the model accurately prescribed preventive maintenance for 80% of the test bearings. The model and data processing flow are relatively simple and compatible with IIoT systems, allowing for low-cost edge inference.