2
an electric motor. Moreover, manual labelling is still a time-
consuming and error-prone task. Unsupervised approaches, on
the other hand, relax this requirement since their operating
principle is the detection of occurrences that deviate from a
normal, i.e., non-faulty, behaviour. Thus, only non-faulty data
are needed to properly design the fault detection algorithm.
These methods will be discussed in more detail in Section II.
The fault detection method proposed in this paper belongs
to the family of data-driven unsupervised approaches. More
in details, motor faults are detected by acquiring vibration
signals and processing them to extract the features necessary
to discriminate faults from non-faults. A deep autoencoder
network is then used for discriminating between normal and
faulty data. Up to the authors’ knowledge, this is the first time
that an unsupervised method based on deep neural networks
is used for detecting motor faults from vibration signals.
The outline of the paper is the following: Section II illus-
trates the recent works recently presented in the literature that
address motor fault detection with knowledge-based methods
and presents the contribution of this paper. Section III de-
scribes in details the proposed method. Section IV presents
the One-Class Support Vector Machine (OC-SVM) approach
that has been compared to the proposed solution. Section V
illustrates the experimental procedure and the obtained results.
Finally, Section VI concludes the paper and presents future
developments.
II. RELATED WORKS AND CONTRIBUTION
As aforementioned, knowledge-based fault detection ap-
proaches can be divided in two classes: supervised and un-
supervised. Among supervised approaches, in [11] the authors
presented a method for analysing the frame vibrations of
a three-phase induction motor during start-up by using the
continuous wavelet transform and Support Vector Machine
(SVM) classifier. Ghate and Dudul [12] measured the stator
current and used a radial-basis-function Multi-layer Perceptron
(MLP) for detecting faults of a three-phase induction motor, in
particular stator winding interturn short and rotor eccentricity
faults. In [25], the authors have used autoencoders for pre-
training a neural network that processes spectral vibration data.
Ince and colleagues [26] applied 1-D Convolutional Neural
Networks (CNNs) directly to the raw motor current signal, thus
eliminating the need for a separate feature extraction stage.
The approach was evaluated for detecting bearing faults and
demonstrated its superiority compared to conventional feature
extraction methods. CNNs have been also applied in [27]
for multiclass classification of bearing faults. As in [26], the
network processes directly raw vibration signals and extracts
features by using wide kernels in the first convolutional layer.
Neural networks have been used also in [5], where the authors
proposed a fault classification method based on Convolutional
Neural Networks that process directly the raw vibration signal.
More in details, the input signal is converted in a 2-D matrix
by stacking zero-padded portions of the time sequence and
then is processed by using the LeNet-5 CNN [28]. The
algorithm proposed in [29] integrates the Support Vector Data
Description framework in the conventional multiclass SVM for
broken rotor bar fault detection in induction motors. The mea-
sured signal is the motor current, from which the stationary
wavelet packet transform is calculated and used as feature.
Sun and colleagues [30] combined compressed sensing and
autoencoder networks for bearing fault diagnosis. In particular,
an autoencoder is trained for extracting features from vibration
signals that then are used by a neural network for supervised
fault classification. A similar approach has been presented
in [31], where a normalized sparse autoencoder is used for
extracting features from raw signals, and a neural network
denoted as local connection network is used to classify the
fault type. Shao and colleagues [32], used an ensemble of
autoencoders with different activation functions for feature
learning and fault diagnosis.
Regarding unsupervised approaches, Soualhi et al. [14] pro-
posed the artificial ant clustering technique to detect faults by
measuring the current and voltage of a squirrel cage induction
motor. As features, they used statistical descriptor such as
standard deviation and mean values of the current, and active
and reactive power signals, spectrum features from the current
signal, power factor and impedance. Cho and colleagues [15]
trained a set of neural networks to model the faults of an
electric motor. One network models the normal behaviour
of the motor, while each additional network models a single
fault condition. Despite each network represents a model of
a normal or faulty system, the approach can be considered
knowledge-based since models are directly obtained from data.
Moreover, models of faulty conditions are used only for fault
isolation, while detection uses only non-faulty data. Razavi-Far
et al. [13] evaluated several one-class classifiers for detecting
broken rotor bar in induction motors. Detection is based on
vibration signals, and for classification they studied density-
based classifiers, nearest neighbour and k-nearest neighbour,
angel-based outlier factor, and k-means. The paper concludes
that k-nearest neighbour is the most performing among the
evaluated approaches.
As evidenced in the examined literature, neural networks
have been widely used in supervised approaches, but scarcely
in unsupervised ones. Approaches based on autoencoders such
as [30]–[32] train them without supervision for extracting
features from the raw vibration signals, but then train the
final classifier in a supervised way. Up to the authors’ knowl-
edge, the only solution is [15], which however is a hybrid
knowledge-based/model-based approach. The main contribu-
tion of this paper is the study of neural networks autoencoders
as a method for detecting motor faults without supervision.
More in details, the approach follows the reconstruction-based
novelty detection paradigm [17], where a model is trained to
reconstruct fault-free (i.e., normal) data with low error and
faulty data with higher error. The magnitude of the error
is then used to detect whether the input data belongs to a
faulty motor or not. For reconstructing the input data, in this
paper we use deep neural networks autoencoders, i.e., neural
networks trained to reconstruct their inputs. Three different
types of autoencoders are evaluated: the MLP autoencoder, the
CNN autoencoder, and a recurrent autoencoder composed of
Long Short-Term Memory (LSTM) units. The autoencoders
take as is input Log-Mel coefficients, which are extracted