Zurich Open Repository and Archive University of Zurich Main Library Strickhofstrasse 39 CH-8057 Zurich www.zora.uzh.ch Year: 2014 Automated fault detection using deep belief networks for the quality inspection of electromotors Sun, Jianwen ; Wyss, Reto ; Steinecker, Alexander ; Glocker, Philipp Abstract: Vibration inspection of electro-mechanical components and systems is an important tool for automated reliable online as well as post-process production quality assurance. Considering that bad electromotor samples are very rare in the production line, we propose a novel automated fault detection method named ”Tilear”, based on Deep Belief Networks (DBNs) training only with good electromotor samples. Tilear consctructs an auto-encoder with DBNs, aiming to reconstruct the inputs as closely as possible. Tilear is structured in two parts: training and decision-making. During training, Tilear is trained only with informative features extracted from preprocessed vibration signals of good electro- motors, which enables the trained Tilear only to know how to reconstruct good electromotor vibration signal features. In the decision-making part, comparing the recorded signal from test electromotor and the Tilear reconstructed signal, allows to measure how well a recording from a test electromotor matches the Tilear model learned from good electromotors. A reliable decision can be made DOI: https://doi.org/10.1515/teme-2014-1006 Other titles: Automatische Fehlerdetektion mittels Deep Belief Netzwerken zur Qualitätskontrolle von Elektromotoren Posted at the Zurich Open Repository and Archive, University of Zurich ZORA URL: https://doi.org/10.5167/uzh-155422 Journal Article Published Version Originally published at: Sun, Jianwen; Wyss, Reto; Steinecker, Alexander; Glocker, Philipp (2014). Automated fault detection using deep belief networks for the quality inspection of electromotors. tm - Technisches Messen, 81(5):255- 263. DOI: https://doi.org/10.1515/teme-2014-1006 DE GRUYTER OLDENBOURG DOI 10.1515/teme-2014-1006 | tm – Technisches Messen 2014; 81(5): 255–263 Beiträge Jianwen Sun*, Reto Wyss, Alexander Steinecker, and Philipp Glocker Automated fault detection using deep belief networks for the quality inspection of electromotors Automatische Fehlerdetektion mittels Deep Belief Netzwerken zur Qualitätskontrolle von Elektromotoren Abstract: Vibration inspection of electro-mechanical com- Ziel, die Eingangssignale so genau wie möglich zu rekon- ponents and systems is an important tool for automated struieren. Tilear besteht aus zwei Teilen: (i) Training und reliable online as well as post-process production qual- (ii) Entscheidung. In der Trainingsphase wird Tilear nur ity assurance. Considering that bad electromotor samples mit Vibrationssignalen guter Motoren angelernt. Auf die- are very rare in the production line, we propose a novel se Weise kann Tilear ausschliesslich Signalmuster rekon- automated fault detection method named “Tilear”, based struieren, die sich einem guten Motor zuordnen lassen. In on Deep Belief Networks (DBNs) training only with good der nachfolgenden Entscheidungsphase wird ein aktuel- electromotor samples. Tilear consctructs an auto-encoder les Vibrationsmuster mit der entsprechenden Rekonstruk- with DBNs, aiming to reconstruct the inputs as closely as tion von Tilear verglichen. Auf diese Weise wird die Abwei- possible. Tilear is structured in two parts: training and chung vom idealen, vorab gelernten Motorsignal quantifi- decision-making. During training, Tilear is trained only ziert und kann für eine Entscheidung in der Qualitätskon- with informative features extracted from preprocessed vi- trolle verwendet werden. bration signals of good electromotors, which enables the trained Tilear only to know how to reconstruct good elec- Schlüsselwörter: Elektromotor, Fehlerdetektion, Deep Be- tromotor vibration signal features. In the decision-making lief Netzwerke, Vibrationssignal, zerstörungsfreie Prü- part, comparing the recorded signal from test electromo- fung, Echtzeit Qualitätskontrolle. tor and the Tilear reconstructed signal, allows to measure how well a recording from a test electromotor matches the || *Corresponding Author: Jianwen Sun, CSEM Alpnach/Institute Tilear model learned from good electromotors. A reliable of Neuroinformatics, University/ETH Zürich, Switzerland, decision can be made. e-mail: [email protected] Reto Wyss: ViDi Systems SA, Villaz-St-Pierre, Switzerland Keywords: Electromotor, fault detection, deep belief net- Alexander Steinecker, Philipp Glocker: CSEM Alpnach, Alpnach works, vibration signals, non-desctructive testing, online Dorf, Switzerland quality inspection. Zusammenfassung: Die Analyse von Vibrationssignalen zur Fehelerdetektion elektromechanischer Komponenten 1 Introduction und Systeme stellt ein wichtiges Werkzeug in zuverlässiger und automatischer Qualitätssicherung des Produktions- Electromotors play an important role in nowadays indus- prozesses dar. Davon ausgehend, dass fehlerhafte Elek- trial applications. Quality inspection of electromotors is tromotoren nur einen geringen Anteil einer Charge aus- essential for manufacturers to assure their products leav- machen, schlagen wir einen neuen Inspektionsansatz na- ing the factory timely with expected quality. It is likewise mens „Tilear” vor. Dieser Ansatz basiert auf einem De- critical for users to perform early failure detection to avoid ep Belief Netzwerk (DBN), welches mit unterschiedlichen possible malfunctions. Automated quality inspection has Signalmustern guter Elektromotoren trainiert wurde. Ti- always been a popular research topic for both scientists lear generiert einen Auto-Encoder mittels DBNs mit dem and engineers. 256 | J. Sun et al., Automated fault detection using deep belief networks DE GRUYTER OLDENBOURG Different techniques have been proposed for electro- performance of the proposed method is compared with motor fault detection. These techniques can be classified a state-of-the-art method, Support Vector Machine (SVM). as signal analysis based methods (SAMs), motor dynamic Finally, conclusions and future directions of this work are model based methods (MMs), and knowledge based meth- giveninSection5. ods (KMs) [1]. SAMs directly analyze measured signals, such as the vibration signal, without a need of accurate motor modeling. However, the dependence of output sig- 2 Automated fault detection using nals on input signals is ignored, and it is difficult to repro- duce analysis results [1]. In contrast, MMs, including e.g. DBNs parameter estimation methods [2, 3] and state estimation Anomaly detection (also called “novelty detection” or methods [4, 5], take both input and output signals into ac- “outlier detection”) has been a mature and active field count. But an accurate model for each specific motor type within diverse research areas [15]. There are many other is needed. proposed anomaly detection techniques [16], such as clas- Recently, methods based on artificial intelligence al- sification based anomaly detection techniques or statis- gorithms, a subdomain of the KMs [1], have been widely tical anomaly detection techniques. DBN is selected be- studied. Most of these readily available techniques are cause of its strong abilities to model high-dimensional on the basis of discriminative learning model. A certain data and to reconstruct the input signal as closely as pos- amount of fault samples are required to perform the fault sible [17, 18]. type classification [6–8]. However, in practical applica- tions, for a well-designed product model, it is extremely difficult to get fault samples in abundance for discrim- 2.1 Deep belief networks inative learning purpose. What makes matters worse is that even a single type of defect typically has many dif- DBN is a probabilistic generative model, which employs ferent sensory manifestations. Alternatively, some previ- a hierarchical structure constructed by stacking Restricted ous studies [9–11] treated the fault detection problem as Boltzmann Machines (RBMs) [17, 19]. As shown in Figure 1, an anomaly detection problem. The core of anomaly detec- RBM contains two layers of neurons: a binary visible layer tion is to recognize the inputs that differ from those under and a binary hidden layer. are the symmetric weights normal conditions. Thus, it is possible to perform fault de- 푊 between visible units v and hidden푠 units . and are bi- tections without the need of collecting a large amount of ℎ 푐 푏 ases to and respectively. Each unit is fully connected to failure data. 푣 ℎ units in the other layer, but there is no connection between Being a very popular research topic in machine learn- units in the same layer. ing society recently, Deep Belief Network’s (DBN) genera- An energy function is used in RBM to model the joint tive nature enables itself a strong feature learning ability. configuration between visible units and hidden units It has shown promise in many tasks, such as hand writ- 푣 . The energy of all possible joint configurations ( in ten digit recognition [12] and speech recognition [13]. In- ℎ 푣 ℎ aRBMisgivenby , ) dustrial applications based on DBN have also appeared, e.g. CSEM’s quality inspection application of complex sur- 푣 ℎ ℎ푇푊푣 푏푇ℎ 푐푇푣 (1) faces [14]. 퐸 ( , ) =− − − The objective of this work is to develop a new auto- The probability of every possible joint configuration푣 ( ℎ mated fault detection system for electromotor quality in- in this energy-based model is , ) spection. Treating the fault detection as an anomaly de- − (2) tection problem,