An adaptive data filtering model for remaining useful life estimation
Kalan faydalı ömür tahmini için uyarlanabilir veri filtreleme modeli
- Tez No: 713513
- Danışmanlar: PROF. DR. JEFF JONES
- Tez Türü: Doktora
- Konular: Sivil Havacılık, Havacılık ve Uzay Mühendisliği, Civil Aviation, Aeronautical Engineering
- Anahtar Kelimeler: Belirtilmemiş.
- Yıl: 2018
- Dil: İngilizce
- Üniversite: Unıversıty Of Warwıck
- Enstitü: Yurtdışı Enstitü
- Ana Bilim Dalı: Havacılık ve Uzay Mühendisliği Ana Bilim Dalı
- Bilim Dalı: Belirtilmemiş.
- Sayfa Sayısı: Belirtilmemiş.
Özet
The field of Prognostics and Health Management is becoming ever more important in the modern maintenance era, with advanced techniques of automation and mechanisation becoming increasingly prevalent. Prognostic technology has promising abilities to forecast remaining useful life and likely future circumstances of complex systems. However, the evolution of data processing and its critical impact on remaining useful life predictions continually demand increasing development so as to meet higher performance levels. There is often a gap between the adequacy of prognostic pre-processing and the prediction methods. One way to reduce this gap would be to design an adaptive data processing method that can filter multidimensional condition monitoring data by incorporating useful information from multiple data sources. Due to the incomplete understanding on the multi-dimensional failure mechanisms and the collaboration between data sources, current prognostic methods lack the ability to deal effectively with complicated interdependency, multidimensional condition monitoring information and noisy data. Further conventional methods are unable to deal with these efficiently. The methodology proposed in this research handles these deficiencies by introducing a prognostic framework that allows the effective use of monitoring data from different resources to predict the lifetime of systems. The methodology presents a feed-forward neural network filtering approach for trajectory similarity-based remaining useful life predictions. The extraction of health indicators is applied as a type of dynamic filtering, in which the time series having full operational conditions are used to train a neural network mapping between raw training inputs and a health indicator output. This trained network function is evaluated by repeating condition monitoring information from multiple data subsets. After the network filtering, the training trajectories are used as baselines to predict the future behaviours of test trajectories. The similarity between these data subsets compares the relationship between the historical performance deterioration of a system's prior operating period with a similar system's degradation behaviour. The proposed prognostic technique, together with dynamic data filtering and remaining useful estimation, holds the promise of increased prediction performance levels. The presented methodology was tested using the PHM08 data challenge provided by the Prognostics Centre of Excellence at NASA Ames Research Centre, and it achieved the overall leading score in the published literature.
Özet (Çeviri)
The field of Prognostics and Health Management is becoming ever more important in the modern maintenance era, with advanced techniques of automation and mechanisation becoming increasingly prevalent. Prognostic technology has promising abilities to forecast remaining useful life and likely future circumstances of complex systems. However, the evolution of data processing and its critical impact on remaining useful life predictions continually demand increasing development so as to meet higher performance levels. There is often a gap between the adequacy of prognostic pre-processing and the prediction methods. One way to reduce this gap would be to design an adaptive data processing method that can filter multidimensional condition monitoring data by incorporating useful information from multiple data sources. Due to the incomplete understanding on the multi-dimensional failure mechanisms and the collaboration between data sources, current prognostic methods lack the ability to deal effectively with complicated interdependency, multidimensional condition monitoring information and noisy data. Further conventional methods are unable to deal with these efficiently. The methodology proposed in this research handles these deficiencies by introducing a prognostic framework that allows the effective use of monitoring data from different resources to predict the lifetime of systems. The methodology presents a feed-forward neural network filtering approach for trajectory similarity-based remaining useful life predictions. The extraction of health indicators is applied as a type of dynamic filtering, in which the time series having full operational conditions are used to train a neural network mapping between raw training inputs and a health indicator output. This trained network function is evaluated by repeating condition monitoring information from multiple data subsets. After the network filtering, the training trajectories are used as baselines to predict the future behaviours of test trajectories. The similarity between these data subsets compares the relationship between the historical performance deterioration of a system's prior operating period with a similar system's degradation behaviour. The proposed prognostic technique, together with dynamic data filtering and remaining useful estimation, holds the promise of increased prediction performance levels. The presented methodology was tested using the PHM08 data challenge provided by the Prognostics Centre of Excellence at NASA Ames Research Centre, and it achieved the overall leading score in the published literature.
Benzer Tezler
- Classification, prediction and state estimation algorithms for predictive maintenance
Kestirimci bakım için sınıflandırma, tahmin ve durum kestirimi algoritmaları
UĞUR YILDIRIM
Doktora
İngilizce
2026
Elektrik ve Elektronik MühendisliğiAdana Alparslan Türkeş Bilim ve Teknoloji ÜniversitesiElektrik-Elektronik Mühendisliği Ana Bilim Dalı
DOÇ. DR. HÜSEYİN AFŞER
- Hava LiDAR nokta bulutundan köprülerin ve üst geçitlerin otomatik çıkarımı
Automatic extraction of bridges and footbridges from airborne LiDAR point cloud
MURAT ERSİN KORKMAZ
Yüksek Lisans
Türkçe
2019
Jeodezi ve FotogrametriKonya Teknik ÜniversitesiHarita Mühendisliği Ana Bilim Dalı
DOÇ. DR. CİHAN ALTUNTAŞ
- Siber güvenlik sistemleri için dinamik ve artımlı makine öğrenmesi yaklaşımları
Dynamic and incremental machine learning approaches for cyber security systems
ENGİN BAYSAL
Doktora
Türkçe
2025
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve KontrolSakarya ÜniversitesiBilgisayar Mühendisliği Ana Bilim Dalı
PROF. DR. CÜNEYT BAYILMIŞ
- Sayısal işaret işlemci ile V32 modem
V32 modem with digital signal processor
ÖMER DOĞAN
Yüksek Lisans
Türkçe
1992
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrolİstanbul Teknik ÜniversitesiDOÇ. DR. MEHMET BÜLENT ÖRENCİK
- An ambient semantic intelligence model for scientific research
Bilimsel araştırma için çevreleyen anlamsal zeka modeli
AHMET SUBAŞI
Doktora
İngilizce
2016
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve KontrolOrta Doğu Teknik ÜniversitesiBilişsel Bilim Ana Bilim Dalı (disiplinlerarası)
PROF. DR. HÜSEYİN CEM BOZŞAHİN
PROF. DR. DENİZ ZEYREK BOZŞAHİN