Machine Faults Detection with Nonlinear Features for Predictive Maintenance Applications

Demirbaş A. M., Yılmaz D.

V. International Applied Statistics Congress (UYIK 2024), İstanbul, Türkiye, 21 - 23 Mayıs 2024, ss.419-429

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Basıldığı Şehir: İstanbul
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.419-429
  • Gazi Üniversitesi Adresli: Evet

Özet

The increasing need for industrial mechanization and diversity of machinery, creating accurate maintenance strategies involving smart technologies has become a significant requirement to ensure continuity in production or service. With predictive maintenance, it is possible to increase efficiency by avoiding additional costs and time losses resulting from traditional maintenance methods or repairs made after problems occur in the machine. The aim of this study is to propose an approach aimed at detecting any maintenance-requiring faults on equipment used with the predictive maintenance concept before the device breaks down. Many studies in the literature had shown that physical measurements such as vibration, sound, and temperature provide information about the machine's fault condition, especially for an electric motor. In this study, the MAFAULDA dataset containing three-axis vibration data from two acceleration sensors at different RPM values was used to examine normal conditions of an electric motor as well as scenarios with horizontal misalignment, imbalance, bearing faults, and vertical misalignment faults. In this study, in addition to the features calculated in the time and frequency domains of acceleration data, some features that have not been used before for this data were used based on nonlinear analysis. These features include average mutual information, Higuchi dimension, Katz dimension, and entropy. Features obtained from both sensors were used for multi-class classification for 10 different fault classes using the MATLAB Classification Learner application, and an accuracy of 99.1% was achieved with the Decision Trees classifier. Then, the classification was made by reducing the number of sensors to one and a 96.1% accuracy rate was achieved with the Decision Trees classifier. When similar studies in the literature are examined, this study achieved the highest classification accuracy achieved with a single sensor. The results obtained in the study are important as they show that nonlinear features provide information that supports the use of fewer sensors.