A hybrid deep learning-based approach for rolling bearing fault prognostics

Abstract

Predictive Maintenance (PdM) has the potential to revolutionize the industry by providing advanced techniques to assess the condition of an industrial system and yield key information that can help optimize maintenance planning and prevent unexpected faults and breakdowns. Nevertheless, PdM is far from being universally applied and it is still the subject of increasing research. Thus, developing new approaches has great relevance to help PdM become a practical reality for the industry. PdM can also bring benefits in terms of sustainability, by reducing human and material resources waste, which is one of the main objectives of Circular Manufacturing initiatives. In this context, rolling bearings are one of the most studied components, as most industrial systems with rotating mechanisms contain bearings, which are prone to a number of faults caused by natural and unnatural wear. In this work, an hybrid Deep Learning (DL) approach is proposed, combining a Convolutional Neural Network (CNN) with a Gated Recurrent Unit (GRU) network to predict Remaining Useful Life (RUL) using rolling bearing vibration data preprocessed with the Short-Time Fourier Transform (STFT). This model was trained and validated using the PRONOSTIA public dataset, which is a popular benchmark for rolling bearing prognostics. The obtained results are satisfactory, providing RUL estimates close to the true values in most test cases, proving the competitiveness of the approach and its potential.