A Sensor Fusion Method using Deep Transfer Learning for Fault Detection in Equipment Condition Monitoring

ÇİNAR E.

16th International Conference on INnovations in Intelligent SysTems and Applications, INISTA 2022, Biarritz, Fransa, 8 - 12 Ağustos 2022 identifier

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Doi Numarası: 10.1109/inista55318.2022.9894141
  • Basıldığı Şehir: Biarritz
  • Basıldığı Ülke: Fransa
  • Anahtar Kelimeler: Deep Learning, Fault Detection and Diagnosis, Sensor Fusion, Smart Manufacturing, Transfer Learning
  • Eskişehir Osmangazi Üniversitesi Adresli: Evet

Özet

© 2022 IEEE.Data-centric fault detection methods utilizing Deep Learning (DL) approaches have recently gained much attention. Researchers have been proposing new sensor fusion methods to increase the accuracy of fault detection and diagnosis. The challenge still lies in designing effective ways of fusing the information from multiple measurement sources and processing them to increase the method's robustness while maintaining its simplicity. This work proposes a new data-level sensor fusion method, which can be expanded into different multi-sensory measurement types and adapted to various applications. Experimental data collected via a data pipeline from a testbed is successfully stored in a database. Utilizing the stored historical time-series data, multi-sensory time-frequency images are overlayed by employing a maximum operation in pixel values of each sensor's spectrogram images. A pre-trained Squeezenet DL model is utilized to leverage the power advantages of Transfer Learning (TL). The method's effectiveness is demonstrated with an induction motor's fault detection use-case having an end-to-end data pipeline solution for equipment monitoring. Higher classification accuracies are observed with respect to single sensor data when TL is employed. Furthermore, the Squeezenet model without TL has failed to demonstrate satisfactory training accuracies with the same dataset.