ïëëïóíîðí ø½÷ îðïç ×ÛÛÛò л®±²¿´ «» · °»®³·¬¬»¼ô ¾«¬ ®»°«¾´·½¿¬·±²ñ®»¼·¬®·¾«¬·±² ®»¯«·®» ×ÛÛÛ °»®³··±²ò Í»» ¸¬¬°æññ©©©ò·»»»ò±®¹ñ°«¾´·½¿¬·±²Á¬¿²¼¿®¼ñ°«¾´·½¿¬·±²ñ®·¹¸¬ñ·²¼»¨ò¸¬³´ º±® ³±®» ·²º±®³¿¬·±²ò
̸· ¿®¬·½´» ¸¿ ¾»»² ¿½½»°¬»¼ º±® °«¾´·½¿¬·±² ·² ¿ º«¬«®» ·«» ±º ¬¸· ¶±«®²¿´ô ¾«¬ ¸¿ ²±¬ ¾»»² º«´´§ »¼·¬»¼ò ݱ²¬»²¬ ³¿§ ½¸¿²¹» °®·±® ¬± º·²¿´ °«¾´·½¿¬·±²ò Ý·¬¿¬·±² ·²º±®³¿¬·±²æ ÜÑ× ïðòïïðçñÌ××òîðïçòîçìíèçèô ×ÛÛÛ
Ì®¿²¿½¬·±² ±² ײ¼«¬®·¿´ ײº±®³¿¬·½
Abstract—This paper develops new deep learning methods,
namely, deep residual shrinkage networks, to improve the feature
learning ability from highly noised vibration signals and achieve a
high fault diagnosing accuracy. Soft thresholding is inserted as
nonlinear transformation layers into the deep architectures to
eliminate unimportant features. Moreover, considering that it is
generally challenging to set proper values for the thresholds, the
developed deep residual shrinkage networks integrate a few
specialized neural networks as trainable modules to automatically
determine the thresholds, so that professional expertise on signal
processing is not required. The efficacy of the developed methods
is validated through experiments with various types of noise.
Index Terms—Deep learning, deep residual networks, fault
diagnosis, soft thresholding, vibration signal.
I. I
NTRODUCTION
OTATING machines are integral to manufacturing, power
supply, transportation, and aerospace industries. However,
because these rotating machines operate under harsh working
environments, failures unavoidably occur in their mechanical
transmission systems and can result in accidents and economic
losses. Accurate fault diagnosis of mechanical transmission
systems can be used to schedule maintenance and extend
service time as well as ensure human safety [1]-[3].
The existing fault diagnosis algorithms for mechanical
transmission systems can be classified into two categories, i.e.,
signal analysis-based methods and machine learning-powered
methods [4]. In general, signal analysis-based fault diagnosis
methods identify faults by detecting the fault-related vibration
components or characteristic frequencies. However, for large
rotating machines, the vibration signals are often composed of
many different vibration components, including meshing of
gears and rotation of shafts and bearings. Further, when faults
are in their early stage, the fault-related components are often
Manuscript received July 16, 2019; revised September 4, 2019; accepted
September 21, 2019. This work was supported by the Key National Natural
Science Foundation of China under Grant U1533202. Paper no. TII-19-3183.
(Corresponding author: Minghang Zhao)
M. Zhao, S. Zhong, and X. Fu are with the School of Naval Architecture and
Ocean Engineering, Harbin Institute of Technology at Weihai, Weihai 264209,
China (e-mail: zhaomh@hit.edu.cn; zhongss@hit.edu.cn;
fuxuyun@hit.edu.cn).
B. Tang is with the State Key Laboratory of Mechanical Transmission,
Chongqing University, Chongqing 400044, China (e-mail:
bptang@cqu.edu.cn).
M. Pecht is with the Center for Advanced Life Cycle Engineering,
University of Maryland, College Park, MD 20742, USA (e-mail:
pecht@umd.edu).
weak and can easily be overwhelmed by other vibration
components and harmonics. As a result, it is often difficult for
the traditional signal analysis-based fault diagnosis methods to
identify the fault-related vibration components and
characteristics frequencies.
Machine learning-powered fault diagnosis methods, on the
other hand, are able to diagnose faults without identifying the
fault-related components and characteristics frequencies. A
number of statistical parameters (e.g., kurtosis, root mean
square, energy, and entropy) can be extracted to represent the
health states, and then a classifier (e.g., multi-class support
vector machines, one-hidden-layer neural networks, and naïve
Bayes classifier) can be trained to diagnose the faults. However,
the extracted statistical parameters are often not discriminative
enough to distinguish the faults, which can lead to low
diagnostic accuracy. As a consequence, finding a
discriminative feature set has become a long-standing
challenge for machine learning-powered fault diagnosis [5].
In recent years, deep learning methods [6], which refer to the
machine learning methods with multiple levels of nonlinear
transformations, have become a useful tool in vibration-based
fault diagnosis. To replace the traditional statistical parameters,
deep learning methods automatically learn features from raw
vibration signals, which can yield higher diagnostic accuracy.
A variety of deep learning methods have been used in machine
fault diagnosis [7]-[14]. For example, Ince et al. [7] employed a
1-dimensional convolutional neural network (ConvNet) to
learn features from current signals for real-time motor fault
diagnosis. Shao et al. [9] applied a convolutional deep belief
network for fault diagnosis of electric locomotive bearings.
However, parameter optimization is often a difficult task for
traditional deep learning methods. The gradients of the error
function, which have to be back-propagated layer by layer,
gradually become inaccurate after flowing through a number of
layers. As a result, the trainable parameters in the beginning
layers (i.e., the layers close to the input layer) cannot be
optimized effectively.
Deep residual networks (ResNets) are an attractive variant of
ConvNets, which use identity shortcuts to ease the difficulty of
parameter optimization [15]. In ResNets, the gradients not only
are back-propagated layer by layer, but also directly flow back
to the beginning layers through identity shortcuts [16]. Due to
the improved parameter optimization, ResNets have been
applied for fault diagnosis in a few recent papers [17]-[20]. For
example, Ma at al. [17] used a ResNet with demodulated
time-frequency features to diagnose planetary gearboxes under
Deep Residual Shrinkage Networks for Fault
Diagnosis
Minghang Zhao, Shisheng Zhong, Xuyun Fu, Baoping Tang, and Michael Pecht, Fellow Member, IEEE