Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Unsupervised feature learning applied to condition monitoring
Luleå University of Technology, Department of Computer Science, Electrical and Space Engineering. (EISLAB)
2017 (English)Doctoral thesis, comprehensive summary (Other academic)
Abstract [en]

Improving the reliability and efficiency of rotating machinery are central problems in many application domains, such as energy production and transportation. This requires efficient condition monitoring methods, including analytics needed to predict and detect faults and manage the high volume and velocity of data. Rolling element bearings are essential components of rotating machines, which are particularly important to monitor due to the high requirements on the operational conditions. Bearings are also located near the rotating parts of the machines and thereby the signal sources that characterize faults and abnormal operational conditions. Thus, bearings with embedded sensing, analysis and communication capabilities are developed.

 

However, the analysis of signals from bearings and the surrounding components is a challenging problem due to the high variability and complexity of the systems. For example, machines evolve over time due to wear and maintenance, and the operational conditions typically also vary over time. Furthermore, the variety of fault signatures and failure mechanisms makes it difficult to derive generally useful and accurate models, which enable early detection of faults at reasonable cost. Therefore, investigations of machine learning methods that avoid some of these difficulties by automated on-line adaptation of the signal model are motivated. In particular, can unsupervised feature learning methods be used to automatically derive useful information about the state and operational conditions of a rotating machine? What additional methods are needed to recognize normal operational conditions and detect abnormal conditions, for example in terms of learned features or changes of model parameters?

 

Condition monitoring systems are typically based on condition indicators that are pre-defined by experts, such as the amplitudes in certain frequency bands of a vibration signal, or the temperature of a bearing. Condition indicators are used to define alarms in terms of thresholds; when the indicator is above (or below) the threshold, an alarm indicating a fault condition is generated, without further information about the root cause of the fault. Similarly, machine learning methods and labeled datasets are used to train classifiers that can be used for the detection of faults. The accuracy and reliability of such condition monitoring methods depends on the type of condition indicators used and the data considered when determining the model parameters. Hence, this approach can be challenging to apply in the field where machines and sensor systems are different and change over time, and parameters have different meaning depending on the conditions. Adaptation of the model parameters to each condition monitoring application and operational condition is also difficult due to the need for labeled training data representing all relevant conditions, and the high cost of manual configuration. Therefore, neither of these solutions is viable in general.

 

In this thesis I investigate unsupervised methods for feature learning and anomaly detection, which can operate online without pre-training with labeled datasets. Concepts and methods for validation of normal operational conditions and detection of abnormal operational conditions based on automatically learned features are proposed and studied. In particular, dictionary learning is applied to vibration and acoustic emission signals obtained from laboratory experiments and condition monitoring systems. The methodology is based on the assumption that signals can be described as a linear superposition of noise and learned atomic waveforms of arbitrary shape, amplitude and position. Greedy sparse coding algorithms and probabilistic gradient methods are used to learn dictionaries of atomic waveforms enabling sparse representation of the vibration and acoustic emission signals. As a result, the model can adapt automatically to different machine configurations, and environmental and operational conditions with a minimum of initial configuration. In addition, sparse coding results in reduced data rates that can simplify the processing and communication of information in resource-constrained systems.

 

Measures that can be used to detect anomalies in a rotating machine are introduced and studied, like the dictionary distance between an online propagated dictionary and a set of dictionaries learned when the machine is known to operate in healthy conditions. In addition, the possibility to generalize a dictionary learned from the vibration signal in one machine to another similar machine is studied in the case of wind turbines.

 

The main contributions of this thesis are the extension of unsupervised dictionary learning to condition monitoring for anomaly detection purposes, and the related case studies demonstrating that the learned features can be used to obtain information about the condition. The cases studies include vibration signals from controlled ball bearing experiments and wind turbines; and acoustic emission signals from controlled tensile strength tests and bearing contamination experiments. It is found that the dictionary distance between an online propagated dictionary and a baseline dictionary trained in healthy conditions can increase up to three times when a fault appears, without reference to kinematic information like defect frequencies. Furthermore, it is found that in the presence of a bearing defect, impulse-like waveforms with center frequencies that are about two times higher than in the healthy condition are learned. In the case of acoustic emission analysis, it is shown that the representations of signals of different strain stages of stainless steel appear as distinct clusters. Furthermore, the repetition rates of learned acoustic emission waveforms are found to be markedly different for a bearing with and without particles in the lubricant, especially at high rotational speed above 1000 rpm, where particle contaminants are difficult to detect using conventional methods. Different hyperparameters are investigated and it is found that the model is useful for anomaly detection with as little as 2.5 % preserved coefficients.

Place, publisher, year, edition, pages
Luleå University of Technology, 2017.
Series
Doctoral thesis / Luleå University of Technology 1 jan 1997 → …, ISSN 1402-1544
National Category
Other Electrical Engineering, Electronic Engineering, Information Engineering
Research subject
Industrial Electronics
Identifiers
URN: urn:nbn:se:ltu:diva-63113ISBN: 978-91-7583-896-0 (print)ISBN: 978-91-7583-897-7 (electronic)OAI: oai:DiVA.org:ltu-63113DiVA: diva2:1090081
Public defence
2017-06-19, A109, Luleå, 10:00 (English)
Opponent
Supervisors
Available from: 2017-04-24 Created: 2017-04-21 Last updated: 2017-05-26Bibliographically approved
List of papers
1. FPGA prototype of machine learning analog-to-feature converter for event-based succinct representation of signals
Open this publication in new window or tab >>FPGA prototype of machine learning analog-to-feature converter for event-based succinct representation of signals
Show others...
2013 (English)In: IEEE International Workshop on Machine Learning for Signal Processing, Piscataway, NJ: IEEE Signal Processing Society, 2013, 6661996Conference paper, Published paper (Refereed)
Abstract [en]

Sparse signal models with learned dictionaries of morphological features provide efficient codes in a variety of applications. Such models can be useful to reduce sensor data rates and simplify the communication, processing and analysis of information, provided that the algorithm can be realized in an efficient way and that the signal allows for sparse coding. In this paper we outline an FPGA prototype of a general purpose "analog-to-feature converter", which learns an overcomplete dictionary of features from the input signal using matching pursuit and a form of Hebbian learning. The resulting code is sparse, event-based and suitable for analysis with parallel and neuromorphic processors. We present results of two case studies. The first case is a blind source separation problem where features are learned from an artificial signal with known features. We demonstrate that the learned features are qualitatively consistent with the true features. In the second case, features are learned from ball-bearing vibration data. We find that vibration signals from bearings with faults have characteristic features and codes, and that the event-based code enable a reduction of the data rate by at least one order of magnitude.

Place, publisher, year, edition, pages
Piscataway, NJ: IEEE Signal Processing Society, 2013
Series
Machine Learning for Signal Processing, ISSN 1551-2541
National Category
Other Electrical Engineering, Electronic Engineering, Information Engineering
Research subject
Industrial Electronics
Identifiers
urn:nbn:se:ltu:diva-31288 (URN)10.1109/MLSP.2013.6661996 (DOI)84893258531 (Scopus ID)56d03ae7-8338-43b1-9be7-7d7862cdba03 (Local ID)56d03ae7-8338-43b1-9be7-7d7862cdba03 (Archive number)56d03ae7-8338-43b1-9be7-7d7862cdba03 (OAI)
Conference
IEEE International Workshop on Machine Learning for Signal Processing : 22/09/2013 - 25/09/2013
Note
Godkänd; 2013; Bibliografisk uppgift: ISSN 1551-2541; 20130805 (fresan)Available from: 2016-09-30 Created: 2016-09-30 Last updated: 2017-10-19Bibliographically approved
2. Unsupervised feature learning for rotating machinery
Open this publication in new window or tab >>Unsupervised feature learning for rotating machinery
(English)In: International Journal of COMADEM, ISSN 1363-7681Article in journal (Other academic) Submitted
Abstract [en]

A smart sensorized bearing can be described as a bearing with built-in sensors for condition monitoring. These bearings require new methods for the processing of the information coming from the sensors. Smart bearings are expected to become the central components in the condition monitoring sensor system of future rotating machines. Condition monitoring typically requires expert knowledge about the machine that is monitored, making it costly to adapt the methods to the different machines, environment and operational conditions. This approach deals with an unsupervised learning method that allows for automatic characterization of signals with repeating structure in the time domain. The method is sparse coding with dictionary learning. We present the information obtained from time domain and frequency domain techniques and describe the conditions required to make the fault diagnosis possible. In contrast, we describe how our approach can autonomously depict deviations from the normal state of operation of machine by monitoring a dictionary of atomic waveforms learned from a signal. We study the propagation over time of a learned dictionary when the vibration of a rotating machine is monitored in normal and faulty states of operation, and we find that the adaptation rates of some atomic waveforms change significantly when a fault occurs.

National Category
Other Engineering and Technologies
Research subject
Industrial Electronics
Identifiers
urn:nbn:se:ltu:diva-63109 (URN)
Available from: 2017-04-21 Created: 2017-04-21 Last updated: 2017-04-21
3. Towards zero-configuration condition monitoring based on dictionary learning
Open this publication in new window or tab >>Towards zero-configuration condition monitoring based on dictionary learning
2015 (English)In: Proceedings of the 23rd European Signal Processing Conference (EUSIPCO 2015): Aug. 31 2015-Sept. 4 2015, Nice, Piscatataway, NJ: IEEE Communications Society, 2015, 1306-1310 p., 7362595Conference paper, Published paper (Refereed)
Abstract [en]

Condition-based predictive maintenance can significantly improve overall equipment effectiveness provided that appropriate monitoring methods are used. Online condition monitoring systems are customized to each type of machine and need to be reconfigured when conditions change, which is costly and requires expert knowledge. Basic feature extraction methods limited to signal distribution functions and spectra are commonly used, making it difficult to automatically analyze and compare machine conditions. In this paper, we investigate the possibility to automate the condition monitoring process by continuously learning a dictionary of optimized shift-invariant feature vectors using a well-known sparse approximation method. We study how the feature vectors learned from a vibration signal evolve over time when a fault develops within a ball bearing of a rotating machine. We quantify the adaptation rate of learned features and find that this quantity changes significantly in the transitions between normal and faulty states of operation of the ball bearing.

Place, publisher, year, edition, pages
Piscatataway, NJ: IEEE Communications Society, 2015
National Category
Other Electrical Engineering, Electronic Engineering, Information Engineering
Research subject
Industrial Electronics
Identifiers
urn:nbn:se:ltu:diva-30747 (URN)10.1109/EUSIPCO.2015.7362595 (DOI)4aa2f985-a6ee-4410-8424-791d788ec410 (Local ID)4aa2f985-a6ee-4410-8424-791d788ec410 (Archive number)4aa2f985-a6ee-4410-8424-791d788ec410 (OAI)
Conference
European Signal Processing Conference : 31/08/2015 - 04/09/2015
Note
Validerad; 2016; Nivå 1; 20150325 (sermar)Available from: 2016-09-30 Created: 2016-09-30 Last updated: 2017-10-20Bibliographically approved
4. Online feature learning for condition monitoring of rotating machinery
Open this publication in new window or tab >>Online feature learning for condition monitoring of rotating machinery
2017 (English)In: Engineering applications of artificial intelligence, ISSN 0952-1976, E-ISSN 1873-6769, Vol. 64, 187-196 p.Article in journal (Refereed) Published
Abstract [en]

Condition-based maintenance of rotating machinery requires efficient condition monitoring methods that enable early detection of abnormal operational conditions and faults. This is a challenging problem because machines are different and change characteristics over time due to wear and maintenance. The efficiency and scalability of conventional condition monitoring methods are limited by the need for manual analysis and re-configuration. The problem to extract relevant features from condition monitoring signals and thereby detect and analyze changes in such signals is a central issue, which in principle can be addressed using machine learning methods. Former work demonstrates that dictionary learning can be used to automatically derive signal features that characterize different operational conditions and faults of a rotating machine, but the use of such methods for online condition monitoring purposes is an open problem. Here we investigate online learning of features using dictionary learning. We describe dictionary distance and signal fidelity based heuristics for anomaly detection, and we study the time--propagated features and sparse approximation of vibration and acoustic emission signals in three different case studies. We present results of numerical experiments with different hyperparameters affecting the approximation accuracy, computational cost, and the adaptation rate of the learned features. We find that the learned features change rapidly when a fault appears in the machine or changes characteristics, and that the dictionary is different in normal and faulty conditions. We find that the learned features change slowly under normal variations of the operational conditions in comparison to the rapid adaptation observed when a fault appears (bearing defects, magnetite particles in the lubricant, or plastic deformation of steel). Furthermore, a sparse signal approximation with 2.5\% preserved coefficients based on a propagated dictionary is sufficient for anomaly detection in the cases considered here. Furthermore, we find that a sparse signal approximation with 2.5\% preserved coefficients based on a propagated dictionary is sufficient for bearing defect detection.

Place, publisher, year, edition, pages
Elsevier, 2017
National Category
Other Electrical Engineering, Electronic Engineering, Information Engineering
Research subject
Industrial Electronics
Identifiers
urn:nbn:se:ltu:diva-63110 (URN)10.1016/j.engappai.2017.06.012 (DOI)2-s2.0-85023608740 (Scopus ID)
Note

Validerad; 2017; Nivå 2; 2017-08-14 (andbra)

Available from: 2017-04-21 Created: 2017-04-21 Last updated: 2017-08-14Bibliographically approved
5. Exploratory Analysis of Acoustic Emissions in Steel using Dictionary Learning
Open this publication in new window or tab >>Exploratory Analysis of Acoustic Emissions in Steel using Dictionary Learning
Show others...
2016 (English)In: IEEE Ultrasonics Symposium 2016, Tours France, September 18-21, 2016, Piscataway, NJ: IEEE conference proceedings, 2016, 7728825Conference paper, Published paper (Refereed)
Abstract [en]

Analysis of acoustic emissions (AE) from steel deformation is a challenging condition monitoring problem due to the high frequencies and data rates involved, and the difficulty to separate signals from noise. The problem to characterize and identify different AE sources calls for methods that goes beyond conventional time and frequency domain analysis. Feature learning is common in the field of machine learning and is successfully used to approximate and classify other kinds of complex signals. Former studies show that AE classification results depend on the choice of predefined features that are extracted from the raw AE signal, but little is known about feature learning in this context. Here we use dictionary learning and sparse coding to optimize a set of shift-invariant features to the AE signal measured in a steel tensile strength test. The specimen undergoes elastic and plastic deformation and eventually cracks. We investigate the learned features and their repetition rates and use principal component analysis (PCA) to illustrate that the resulting sparse AE code is useful for classification of the three strain stages, without reference to the signal amplitude. Therefore, feature learning is a potentially useful approach to the AE analysis problem, which also opens up for further studies of automated methods for anomaly detection in AE.

Place, publisher, year, edition, pages
Piscataway, NJ: IEEE conference proceedings, 2016
Series
Proceedings - IEEE Ultrasonics Symposium, ISSN 1948-5719
Keyword
Dictionary Learning, Acoustic Emission
National Category
Other Electrical Engineering, Electronic Engineering, Information Engineering
Research subject
Industrial Electronics; Machine Elements
Identifiers
urn:nbn:se:ltu:diva-59776 (URN)10.1109/ULTSYM.2016.7728825 (DOI)000387497400452 ()2-s2.0-84996567218 (Scopus ID)978-1-4673-9897-8 (ISBN)978-1-4673-9898-5 (ISBN)
Conference
IEEE Ultrasonics Symposium 2016, Tours France, September 18-21, 2016
Available from: 2016-10-16 Created: 2016-10-16 Last updated: 2017-10-19Bibliographically approved
6. Detection of particle contaminants in rolling element bearings with unsupervised acoustic emission feature learning
Open this publication in new window or tab >>Detection of particle contaminants in rolling element bearings with unsupervised acoustic emission feature learning
2017 (English)In: Tribology International, ISSN 0301-679X, E-ISSN 1879-2464Article in journal (Other academic) Submitted
National Category
Tribology Other Electrical Engineering, Electronic Engineering, Information Engineering
Research subject
Machine Elements; Industrial Electronics
Identifiers
urn:nbn:se:ltu:diva-63112 (URN)
Available from: 2017-04-21 Created: 2017-04-21 Last updated: 2017-08-28
7. A dictionary learning approach to monitoring of wind turbine drivetrain bearings
Open this publication in new window or tab >>A dictionary learning approach to monitoring of wind turbine drivetrain bearings
2017 (English)In: Mechanical systems and signal processing, ISSN 0888-3270, E-ISSN 1096-1216Article in journal (Other academic) Submitted
Abstract [en]

Condition monitoring and predictive maintenance are central for efficient operation of wind farms due to the challenging operating conditions, rapid technology development and high number of aging wind turbines. In particular, preventive maintenance planning requires early detection of faults with few false positives. This is a challenging problem due to the complex and weak signatures of some faults, in particular of faults occurring in some of the drivetrain bearings. Here, we investigate recently proposed condition monitoring methods based on unsupervised dictionary learning using vibration data recorded from three wind turbines over about four years of operation, thereby contributing novel test results based on real world data. Results of former studies addressing condition--monitoring tasks using dictionary learning indicate that unsupervised feature learning is useful for diagnosis and anomaly detection purposes. However, these studies are based on data from test rigs operating under controlled conditions. Furthermore, most former studies focus on classification tasks using relatively small sets of labeled data, which are useful for quantitative method comparisons but gives little information about how useful these approaches are in practice. In this study dictionaries are learned from gearbox vibrations in three different turbines known to be in healthy conditions, and the dictionaries are subsequently propagated over a few years of monitoring data when faults are known to occur. We calculate the dictionary distance between the initial and propagated dictionaries and find time periods of abnormal dictionary adaptation starting six months before a drivetrain bearing replacement and one year before the resulting gearbox replacement. When repeating that experiment with a dictionary that initially is learned from the vibration of another type of rotating machine, the corresponding difference of dictionary distances is three times lower and do not appear abnormal. We also investigate the distance between dictionaries learned from geographically nearby turbines of the same type in healthy conditions and find that the features learned are similar, and that a dictionary learned from one turbine can be useful for monitoring of another similar turbine.

National Category
Other Electrical Engineering, Electronic Engineering, Information Engineering Tribology
Research subject
Industrial Electronics; Machine Elements
Identifiers
urn:nbn:se:ltu:diva-63111 (URN)
Available from: 2017-04-21 Created: 2017-04-21 Last updated: 2017-08-28
8. Dictionary Learning with Equiprobable Matching Pursuit
Open this publication in new window or tab >>Dictionary Learning with Equiprobable Matching Pursuit
2017 (English)Conference paper, Published paper (Refereed)
Abstract [en]

Sparse signal representations based on linear combinations of learned atomshave been used to obtain state-of-the-art results in several practical signalprocessing applications. Approximation methods are needed to processhigh-dimensional signals in this way because the problem to calculate optimalatoms for sparse coding is NP-hard. Here we study greedy algorithms forunsupervised learning of dictionaries of shift-invariant atoms and propose anew method where each atom is selected with the same probability on average,which corresponds to the homeostatic regulation of a recurrent convolutionalneural network. Equiprobable selection can be used with several greedyalgorithms for dictionary learning to ensure that all atoms adapt duringtraining and that no particular atom is more likely to take part in the linearcombination on average. We demonstrate via simulation experiments thatdictionary learning with equiprobable selection results in higher entropy ofthe sparse representation and lower reconstruction and denoising errors, bothin the case of ordinary matching pursuit and orthogonal matching pursuit withshift-invariant dictionaries. Furthermore, we show that the computational costsof the matching pursuits are lower with equiprobable selection, leading tofaster and more accurate dictionary learning algorithms.

Place, publisher, year, edition, pages
Piscataway, NJ: Institute of Electrical and Electronics Engineers (IEEE), 2017
Keyword
dictionary learning, sparse approximation, matching pursuit, unsupervised learning, homeostatic regulation, neuromorphic engineering
National Category
Other Electrical Engineering, Electronic Engineering, Information Engineering
Research subject
Industrial Electronics
Identifiers
urn:nbn:se:ltu:diva-60768 (URN)10.1109/IJCNN.2017.7965902 (DOI)978-1-5090-6182-2 (ISBN)
Conference
2017 International Joint Conference on Neural Networks (IJCNN), Anchorage, AK, 14-19 May 2017
Funder
The Kempe Foundations
Available from: 2016-11-29 Created: 2016-11-29 Last updated: 2017-10-19Bibliographically approved

Open Access in DiVA

fulltext(8608 kB)521 downloads
File information
File name FULLTEXT01.pdfFile size 8608 kBChecksum SHA-512
7c7a05ebf1f6dd05ba8c0a3453e07e4f4abbd89a918efb4dab78d05c77770b76313c352cfc3a212ecba05f02ee541ff15cf0eb9c8c736cabcc2fd8b8ba9867b7
Type fulltextMimetype application/pdf

Search in DiVA

By author/editor
Martin del Campo Barraza, Sergio
By organisation
Department of Computer Science, Electrical and Space Engineering
Other Electrical Engineering, Electronic Engineering, Information Engineering

Search outside of DiVA

GoogleGoogle Scholar
Total: 521 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

Total: 1006 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf