T4PdM: a Deep Neural Network based on the Transformer Architecture for Fault Diagnosis of Rotating Machinery
Authors:
Erick Giovani Sperandio Nascimento,
Julian Santana Liang,
Ilan Sousa Figueiredo,
Lilian Lefol Nani Guarieiro
Abstract:
Deep learning and big data algorithms have become widely used in industrial applications to optimize several tasks in many complex systems. Particularly, deep learning model for diagnosing and prognosing machinery health has leveraged predictive maintenance (PdM) to be more accurate and reliable in decision making, in this way avoiding unnecessary interventions, machinery accidents, and environmen…
▽ More
Deep learning and big data algorithms have become widely used in industrial applications to optimize several tasks in many complex systems. Particularly, deep learning model for diagnosing and prognosing machinery health has leveraged predictive maintenance (PdM) to be more accurate and reliable in decision making, in this way avoiding unnecessary interventions, machinery accidents, and environment catastrophes. Recently, Transformer Neural Networks have gained notoriety and have been increasingly the favorite choice for Natural Language Processing (NLP) tasks. Thus, given their recent major achievements in NLP, this paper proposes the development of an automatic fault classifier model for predictive maintenance based on a modified version of the Transformer architecture, namely T4PdM, to identify multiple types of faults in rotating machinery. Experimental results are developed and presented for the MaFaulDa and CWRU databases. T4PdM was able to achieve an overall accuracy of 99.98% and 98% for both datasets, respectively. In addition, the performance of the proposed model is compared to other previously published works. It has demonstrated the superiority of the model in detecting and classifying faults in rotating industrial machinery. Therefore, the proposed Transformer-based model can improve the performance of machinery fault analysis and diagnostic processes and leverage companies to a new era of the Industry 4.0. In addition, this methodology can be adapted to any other task of time series classification.
△ Less
Submitted 7 April, 2022;
originally announced April 2022.
