Al-Tekreeti Watban Khalid Fahmi a , Kazem Reza Kashyzadeh b , Siamak Ghorbani a
الملخص الانجليزي
This research investigates the effectiveness of various vibration data acquisition
techniques coupled with different machine learning models for detecting
anomalies and classifying them. To this end, synthetic vibration data was
generated for techniques such as eddy current proximity transducers (ECPT),
accelerometer sensor, blade tip timing, laser doppler vibrometer (LDV), and
strain gauge. Afterward, the data was pre-processed and used to train gradient
boosting machine, support vector machine, and random forest models.
Performance evaluation metrics, including accuracy, recall, F1-score, receiver
operating characteristic, and area Under curve were employed to assess the
models, revealing varying degrees of success across combining techniques and
models. Notable achievements were observed for the random forest model
coupled with the eddy current proximity transducers technique, underscoring
the significance of informed technical selection and model optimization in
enhancing vibration anomaly detection systems in combined cycle power
plants. The results showed that the LDV technique has a significant increase in
accuracy from about 0.49 to approximately 0.52, while the ECPT technique has
improved from about 0.9 to close 1.0. These advances highlight the growing
accuracy of the methods and enable the development of more efficient and
reliable learning machines.
تاريخ النشر
06/11/2024
الناشر
J. Comp. App. Res. Mech. Eng.
رقم المجلد
رقم العدد
رابط DOI
DOI: 10.22061/jcarme.2024.10797.2415
الصفحات
12
الكلمات المفتاحية
Anomaly detection, Machine learning, Eddy current proximity transducers, Blade tip timing, Laser doppler vibrometer
وطبان خالد فهمي التكريتي | Al- TEKREETI Watban Khalid Fahmi | 3420