Deteksi Dini Kegagalan pada Turbin Angin Menggunakan Long Short Term Memory (LSTM) dengan Attention Mechanism dan Dynamic Threshold Berbasis Health Index

Cahyono, Shafira Melyana (2026) Deteksi Dini Kegagalan pada Turbin Angin Menggunakan Long Short Term Memory (LSTM) dengan Attention Mechanism dan Dynamic Threshold Berbasis Health Index. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5007221028-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy

Abstract

Penelitian ini bertujuan mengembangkan model deteksi dini kerusakan dengan membentuk Health Index (HI) yang merepresentasikan kondisi kesehatan turbin dan meminimalkan false alarm pada kondisi startup. Data berupa sinyal getaran turbin angin Aventa dengan sampling rate 200 Hz. Model yang digunakan Long Short Term Memory (LSTM) dengan attention mechanism untuk bobot adaptif fitur dan waktu serta dynamic threshold sebagai batas deteksi. Preprocessing meliputi data cleaning, outlier handling, segmentasi sinyal, dan detrending, kemudian ekstraksi 15 fitur domain waktu dan frekuensi. Hasil ekstraksi diproses dengan sliding window berukuran 30 dengan step size 10 dan direkonstruksi oleh LSTM. Selisih fitur hasil rekonstruksi dan fitur asli dinyatakan sebagai reconstruction error. Hasil hyperparameter tuning adalah 1 layer, batch size 32, hidden dimension 32, dan parameter EWMA 0,4. Semakin kecil reconstruction error, maka pola window tersebut masih menyerupai pola yang dipelajari dari data normal. Hasil menunjukkan pada fase startup model dengan attention memiliki nilai HI lebih rendah dibandingkan model tanpa attention. Pada fase startup window 1030 hingga 1035, untuk model attention mechanism nilai HI yaitu 0,287, 0,304, 0,319, 0,332, 0,341 dan 0,351. Untuk model tanpa attention 0,299, 0,314, 0,337, 0,351, 0,361 dan 0,371. Selain itu, penggunaan dynamic threshold dapat menyesuaikan tren HI. Performa model dibuktikan dengan nilai accuracy 98,8%, precision 97,9%, recall 97,1%, F1-score 97,5%, robustness 98,1%, False Alarm Rate (FAR) 0,7% dan Missed Alarm Rate (MAR) 2,9%. Sehingga model LSTM attention mechanism dengan dynamic threshold dapat membentuk representasi health index yang sesuai dengan penentuan alarm early detection pada kondisi degradasi dan nilai health index tidak meningkat saat startup.
========================================================================================================================
This study develops an early fault detection model by constructing a Health Index (HI) that represents wind turbine health while reducing false alarms during startup. The dataset vibration signals from an Aventa wind turbine sampled at 200 Hz. The proposed model employs a Long Short Term Memory (LSTM) with an attention mechanism to assign adaptive weights to features and time steps, together with a dynamic threshold for alarm detection. Preprocessing includes data cleaning, outlier handling, signal segmentation, and detrending, followed by the extraction of 15 time and frequency domain features. The extracted features are processed using a sliding window with a window size of 30 and a step size of 10 before being reconstructed by the LSTM. The difference between the reconstructed and original features is defined as the reconstruction error. The optimal hyperparameters are one LSTM layer, a batch size of 32, a hidden dimension of 32, and an EWMA parameter of 0,4. A lower reconstruction error indicates the window pattern similar to the normal operating condition learned by the model. During the startup phase, the attention based model produces lower HI values than the model without attention. For startup windows 1030–1035, the attention model produce HI values of 0.287, 0,304, 0,319, 0,332, 0,341, and 0,351 compared with 0,299, 0,314, 0,337, 0,351, 0,361, 0,371 for the model without attention. In addition, the dynamic threshold effectively adapts to the HI trend. The proposed model achieves an accuracy of 98.8%, precision of 97.9%, recall of 97.1%, F1-score of 97.5%, robustness of 98.1%, a False Alarm Rate (FAR) of 0.7%, and a Missed Alarm Rate (MAR) of 2.9%. These results demonstrate that the proposed LSTM model with an attention mechanism and dynamic threshold provides a representative health index for early fault detection under degradation while preventing unnecessary HI increases during startup.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deteksi, Getaran, Turbin Angin, Machine Learning
Subjects:	T Technology > TJ Mechanical engineering and machinery > TJ828 Wind turbines
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Mechanical Engineering > 21201-(S1) Undergraduate Thesis
Depositing User:	Shafira Melyana Cahyono
Date Deposited:	03 Aug 2026 07:14
Last Modified:	03 Aug 2026 07:14
URI:	http://repository.its.ac.id/id/eprint/142362

Actions (login required)

View Item