Predictive Maintenance (PdM) menggunakan Active dan Semi-Supervised Machine Learning Pada Mesin Industri

Andriani, Adelina Zian (2021) Predictive Maintenance (PdM) menggunakan Active dan Semi-Supervised Machine Learning Pada Mesin Industri. Masters thesis, Institut Teknologi Sepuluh Nopember.

[thumbnail of 02411950010001-Master_Thesis.pdf] Text
02411950010001-Master_Thesis.pdf - Accepted Version
Restricted to Repository staff only until 1 April 2023.

Download (2MB) | Request a copy

Abstract

Sejalan dengan kemajuan Industri 4.0 dengan peluang pemanfaatan sensor dan teknologi Machine Learning (ML), menjadikan praktik Predictive Maintenance (PdM) lebih mudah dilakukan. Untuk menunjang PdM dengan ML, manufaktur perlu menyediakan data yang mendukung proses pembelajaran mesin. Namun, riil nya mayoritas data tidak berlabel dan masih membutuhkan proses pelabelan manual untuk menunjang proses pembelajaran, yang mana berisiko terhadap kesalahan, memakan biaya, dan tenaga. Oleh karena itu, penelitian saat ini menggunakan integrasi Active Learning (AL) dan Semi-Supervised Learning (SSL) dalam mengatasi masalah pelabelan dan optimalisasi ketersediaan data tidak berlabel untuk mendukung pembentukan model PdM dengan tingkat generalisasi yang lebih baik. Algoritme AL membuat expertise hanya melakukan pelabelan pada sampel yang paling informatif, sedangkan sampel tersisa secara otomatis diberi label oleh anotator mesin melalui algoritme SSL. Pertama, rekapan data multi sensor tidak berlabel pada database main server dan sedikit data berlabel yang tersedia, menjadi sampel pembelajaran mesin. Kedua, skema AL memilih sampel tak berlabel yang paling berharga, untuk diberi label dan ditambahkan ke kumpulan data pelatihan. Ketiga, skema SSL untuk mengoptimalisasi penggunaan data, menggunakan sampel tersisa untuk dilabeli. Terakhir, berdasarkan data pelatihan yang terbentuk, model prediktif dilatih untuk menunjang prediksi kelas kegagalan pada fault diagnostic, yang hasilnya akan menunjang prediksi pergerakan nilai sensor pada fault prognostic. Sehubungan dengan pemilihan algoritme ML, hasil pelatihan Random Forest (RF) mampu memprediksi model diganostik 99,85% dan dengan menggunakan Support Vector Regression (SVR) dapat memprediksi model prognostik 97,09%. Hasil prediksi mendukung pembentukan strategi pemeliharaan melalui pembuatan skema fault diagnostic dan prognostic berbasis machine learning, serta perencanaan planned maintenance atas kegagalan peralatan yang terjadi.
====================================================================================================
In line with the advancement of Industry 4.0 with the opportunity to use sensors and Machine Learning (ML) technology, making Predictive Maintenance (PdM) practices easier to do. To support PdM with ML, manufacturers need to provide data that supports the machine learning process. However, in reality the majority of data is unlabeled and still requires a manual labeling process to support the learning process, which is risky, costly, and labor intensive. Therefore, current research uses the integration of Active Learning (AL) and Semi- Supervised Learning (SSL) in overcoming labeling problems and optimizing the availability of unlabeled data to support the formation of PdM models with a better degree of generalization. The AL algorithm allows expertise to only label the most informative samples, while the remaining samples are automatically labeled by machine annotators via the SSL algorithm. First, multi-sensor data recaps are not labeled on the main server database and the little labeled data available becomes a machine learning sample. Second, the AL scheme selects the most valuable unlabelled samples, to label and add to the training data set. Third, the SSL scheme to optimize data usage, using the remaining samples to be labeled. Finally, based on the formed training data, the predictive model is trained to support the prediction of the failure class on the fault diagnostic, the results of which will support the prediction of the movement of the sensor value on the fault prognostic. In connection with the choice of the ML algorithm, the results of Random Forest (RF) training were able to predict the diagnostic model of 99.85% and using Support Vector Regression (SVR) to predict the prognostic model of 97.09%. The prediction results support the formation of a maintenance strategy through making machine learning-based fault diagnostic and prognostic schemes, as well as planned maintenance planning for equipment failures that occur.

Keywords: Predictive Maintenance (PdM), Active Learning, Semi-Supervised Learning

Item Type: Thesis (Masters)
Uncontrolled Keywords: Predictive Maintenance (PdM), Active Learning, Semi-Supervised Learning
Subjects: T Technology > TJ Mechanical engineering and machinery > TJ174 Maintenance and repair of machinery
Divisions: Faculty of Industrial Technology and Systems Engineering (INDSYS) > Industrial Engineering > 26101-(S2) Master Thesis
Depositing User: Adelina Zian Andriani
Date Deposited: 03 Mar 2021 07:20
Last Modified: 03 Mar 2021 07:20
URI: http://repository.its.ac.id/id/eprint/83296

Actions (login required)

View Item View Item