Online automatic diagnosis of wind turbine bearings progressive degradations under real experimental conditions based on unsupervised machine learning

Ali, Jaouher Ben; Saïdi, Lotfi; Harrath, Salma; Bechhoefer, Eric; Benbouzid, Mohamed

doi:10.1016/j.apacoust.2017.11.021

Cited by 107 publications

(42 citation statements)

References 45 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…These data sets also present additional problems. First, all but the GPMS (Ben Ali et al, 2018) data set, were obtained under laboratory conditions, making model evaluation sub-optimal. Second, they consist of a very limited number of instances (Table 4), which potentially introduces problem with data snooping.…”

Section: Resultsmentioning

confidence: 99%

Remaining Useful Life Estimation of Bearings

Ferreira

Sousa

2021

IJPHM

View full text Add to dashboard Cite

In the domain of predictive maintenance, when trying to repli- cate and compare research in remaining useful life estimation (RUL), several inconsistencies and errors were identified in the experimental methodology used by various researchers. This makes the replication and the comparison of results diffi- cult, thus severely hindering both progress in this research do- main and its practical application to industry. We survey the literature to evaluate the experimental procedures that were used, and identify the most common errors and omission in both experimental procedures and reporting. A total of 70 papers on RUL were audited. From this meta- analysis we estimate that approximately 11% of the papers present work that will allow for replication and comparison. Surprisingly, only about 24.3% (17 of the 70 articles) com- pared their results with previous work. Of the remaining work, 41.4% generated and compared several models of their own and, somewhat unsettling, 31.4% of the researchers made no comparison whatsoever. The remaining 2.9% did not use the same data set for comparisons. The results of this study were also aggregated into 3 categories: problem class selec- tion, model fitting best practices and evaluation best practices. We conclude that model evaluation is the most problematic one. The main contribution of the article is a proposal of an ex- perimental protocol and several recommendations that specif- ically target model evaluation. Adherence to this protocol should substantially facilitate the research and application of RUL prediction models. The goals are to promote the collab- oration between scholars and practitioners alike and advance the research in this domain.

show abstract

Section: Resultsmentioning

confidence: 99%

Remaining Useful Life Estimation of Bearings

Ferreira

Sousa

2021

IJPHM

View full text Add to dashboard Cite

show abstract

“…Our main contribution to modify the autoencoder in this work is that the basic ELM theories for training the autoencoder show that: "After training the Autoencoder, we can encode features using the transpose of output weights matrix [13], Equation (14)". However, mathematically it is proven that the best encoding can be obtained by using the inverse matrix of the transpose of output weights as shown in Equation (15).…”

Section: Training Of the Proposed Networkmentioning

confidence: 99%

“…Zhou et al [13] proposed a Stacked ELM (S-ELM) where a stack of small ELMs is specially designed for solving large and complex data problems. Ben Ali et al [14] presents a new unsupervised learning classification tool of extracted data based on the Adaptive Resonance Theory 2 (ART2) for high-speed shaft bearing prognosis. Li et al [15] proposed an improved OS-ELM which is one of the ELM variants with an adaptive forgetting factor and an update selection strategy to predict gas utilization ratio using time-varying data.…”

mentioning

confidence: 99%

Aircraft Engines Remaining Useful Life Prediction with an Improved Online Sequential Extreme Learning Machine

et al. 2020

Self Cite

View full text Add to dashboard Cite

The efficient data investigation for fast and accurate remaining useful life prediction of aircraft engines can be considered as a very important task for maintenance operations. In this context, the key issue is how an appropriate investigation can be conducted for the extraction of important information from data-driven sequences in high dimensional space in order to guarantee a reliable conclusion. In this paper, a new data-driven learning scheme based on an online sequential extreme learning machine algorithm is proposed for remaining useful life prediction. Firstly, a new feature mapping technique based on stacked autoencoders is proposed to enhance features representations through an accurate reconstruction. In addition, to attempt into addressing dynamic programming based on environmental feedback, a new dynamic forgetting function based on the temporal difference of recursive learning is introduced to enhance dynamic tracking ability of newly coming data. Moreover, a new updated selection strategy was developed in order to discard the unwanted data sequences and to ensure the convergence of the training model parameters to their appropriate values. The proposed approach is validated on the C-MAPSS dataset where experimental results confirm that it yields satisfactory accuracy and efficiency of the prediction model compared to other existing methods.

show abstract

“…f1 to f4 listed in Table II are adopted based on this principle. A brief explanation of these features can be found in [31] signal-processing applications. In many cases, the most useful information is hidden in the frequency content of the signal.…”

Section: Feature Extractionmentioning

confidence: 99%

Unobtrusive Screening of Central Sleep Apnea From Pressure Sensors Measurements: A Patient-Specific Longitudinal Study

Azimi

Bouchard

Goubran

et al. 2020

IEEE Trans. Instrum. Meas.

View full text Add to dashboard Cite

Historically, the lack of patients' sleep histories has caused low identification of sleep apnea (SA) and referral rates. Moreover, the costly and time-consuming nature of polysomnography (PSG) as a standard clinical test for detecting SA and the lack of sleep clinics has created a demand for suitable home-based monitoring devices. Pressure measurement using a pressure sensitive mat (PSM) can address the challenges found in current sleep-monitoring solutions. The noncontact PSM has a potential to replace obtrusive breathing sensors in the sleep lab and to be used as a prescreening tool for patients suspected of having SA. Applying classical support vector machine (SVM), this article presents a personalized system based on the measurements of each patient to detect central SA (CSA) events and monitor sleep characteristics longitudinally. For this purpose, sensor set-ups were installed in nine seniors' homes to collect unsupervised pressure data in approximately one year ranging from 8 to 12 months. Cost-based and resampling-based approaches were examined to combat imbalanced data. The results showed that the cost-based method outperformed other methods. Next, the patient-specific system was used to determine the total number of CSA events, as well as their starting time and duration in each day. The SA severity was measured by the central apnea index (CAI). In addition, other sleep characteristics such as bed occupancy (BO), day clock, and night clock were extracted from the PSM measurements. The impact of longitudinal sleep monitoring could be in tracking SA treatment progression, and possibly providing information on the interaction between SA and other disease progressions.

show abstract

Online automatic diagnosis of wind turbine bearings progressive degradations under real experimental conditions based on unsupervised machine learning

Cited by 107 publications

References 45 publications

Remaining Useful Life Estimation of Bearings

Remaining Useful Life Estimation of Bearings

Aircraft Engines Remaining Useful Life Prediction with an Improved Online Sequential Extreme Learning Machine

Unobtrusive Screening of Central Sleep Apnea From Pressure Sensors Measurements: A Patient-Specific Longitudinal Study

Contact Info

Product

Resources

About