Automatic Feature Extraction for Bearings'Degradation Assessment using Minimally Pre-processed Time series and Multi-modal Feature Learning

Oct 28, 2026

Upload by Rhys

Maintenance activities benefit from machine learning to monitor asset health, yet diverse measurable signals such as temperature and vibration and heterogeneous degradation behaviors make manual feature extraction costly and time-consuming. This work proposes data-driven feature learning to transform minimally preprocessed time series from multiple sources into informative, fused representations. Multiple autoencoder architectures are evaluated for vibration- and temperature-based bearings degradation, measuring both recognition accuracy and the quality of learned features on a real-world dataset, and benchmarking against manual features and state-of-the-art methods.

LanguageEnglish

CategoryTechnology

FormatPDF

Pages10

Reading Time25 min

AccessFree to download

1 INTRODUCTION

# Automatic Feature Extraction for Bearings'Degradation Assessmentusing Minimally Pre-processed Time series and Multi-modal FeatureLearning Antonio L.Alfeo¹,2,Mario G.C.A.Cimino¹,²and Guido Gagliardi1,3,4,* ¹Department of Information Engineering,University of Pisa,Pisa,Italy2Bioengineering and Robotics Research Center E.Piaggio,University of Pisa,Pisa,Italy³Department of Information Engineering(DINFO),University of Florence,Florence,Italy4Department of Electrical Engineering,KU Leuven,Leuven,Belgium Keywords: Feature Learning,Data Fusion,Deep Learning,Multimodal Autoencoder,Contrastive Learning,PredictiveMaintenance,Smart Manufacturing. Abstract:Maintenance activities can be better planned by employing machine learning technologies to monitor an asset'shealth conditions.However,the variety of observable measures(e.g.temperature,vibration)and behaviourscharacterizing the health degradation process results in time-consuming manual feature extraction to ensureaccurate degradation stage recognitions.Indeed,approaches able to provide automatic feature extraction frommultiple and heterogeneous sources are more and more required in the field of predictive maintenance.This is-sue can be addressed in a data-driven fashion by using feature learning technology,enabling the transformationof minimally processed time series into informative features.Given its capability of discovering meaningfulpatterns in data while enabling data fusion,many feature learning approaches are based on deep learningtechnology(e.g.autoencoders).In this work,an architecture based on autoencoders is used to automaticallyextract degradation-representative features from minimally preprocessed time series of vibration and temper-ature data.Different autoencoder architectures are implemented to compare different data fusion strategies.The proposed approach is tested considering both the recognition performances and the quality of the learnedfeatures with a publicly available real-world dataset about bearings'progressive degradation.The proposedapproach is also compared against manual feature extraction and the state-of-the-art technology in featurelearning. ## 1 INTRODUCTION RUL predictions are often unreliable since theyare greatly affected by the usage of the assets whilein an unhealthy stage (Lei et al.,2018).For this rea-son,many real-world applications provide degrada-tion stage estimations rather than RUL predictions. The paradigm known as Industry 4.0 recommendsthe integration of machine learning in industrial pro-cesses to improve them(Alfeo et al.,2021)and avoidproduction inefficiencies(Alfeo et al.,2020).Re-garding the maintenance procedures,such technolo-gies are enabling the so-called Predictive Mainte-nance(PdM)(Jimenez et al.,2020).PdM aims atdriving the maintenance operations to execute themjust before the breaking point rather than after a fail-ure or according to some regular schedule.This al-lows avoiding the costs due to a failure while still ex-ploiting the whole remaining useful life (RUL)of anasset's component(Wan et al.,2017). The number of stages characterizing the degra-dation process results from a trade-off between theinterpretability of the prediction and the complex-ity of the degradation process.The more the asset'sdegradation process is consistent and progressive,themore it can be effectively modeled with a few easy-to-interpret stages.Most of the research works(Leiet al.,2018)divide the degradation process into three(Vinh et al.,2009),others into four(Scanlon et al.,2012),or even five stages(Kimotho et al.,2013). However,due to the many possible measures tomonitor an asset's health condition,and the diversity 94 Automatic Feature Extraction for Bearings'Degradation Assessment using Minimally Pre-processed Time series and Multi-modal FeatureLearning of the degradation processes across industries and ma-chines,it is difficult to provide a high-quality featureextraction process that is also generalizable amongdiffe

FAQs

Why is manual feature extraction challenging in predictive maintenance for bearings?

The degradation process can be characterized by many heterogeneous measures, such as temperature and vibration, and varies across industries and machines. This makes it difficult to build a generalizable and high-quality feature extraction pipeline, and it also increases the effort required for manual transformation of raw signals into useful features.

What is the main idea of the proposed approach?

The approach uses feature learning with autoencoder-based architectures to automatically extract degradation-representative features from minimally preprocessed vibration and temperature time series. It aims to support data fusion and discover meaningful patterns without relying on manual feature engineering.

How is the method evaluated in the study?

The architecture is tested on a publicly available real-world dataset of bearings progressive degradation. Evaluation considers both recognition performance and the quality of the learned features, and it is compared against manual feature extraction and state-of-the-art feature learning techniques.

Oct 28, 2026

Upload by Rhys