A novel method of combining nonlinear frequency spectrum and deep
learning for complex system fault diagnosis
Lerui Chen
⇑
, Zerui Zhang, Jianfu Cao, Xiaoqi Wang
State Key Laboratory for Manufacturing Systems Engineering, Xi’an JiaoTong University, Xi’an 710049, China
article info
Article history:
Received 9 July 2019
Received in revised form 9 September 2019
Accepted 19 October 2019
Available online 25 October 2019
Keywords:
Nonlinear frequency spectrum
Output frequency response functions
Stack denoising auto-encoders
Fault diagnosis
abstract
A novel fault diagnosis method is proposed for complex systems by combining nonlinear frequency spec-
trum and stacked denoising auto-encoders (SDAE). In order to solve the problem of large calculation
amount of generalized frequency response functions (GFRF), one-dimensional nonlinear output fre-
quency response functions (NOFRF) are used to obtain nonlinear frequency spectrum. In order to over-
come the problem of weak ability of fault features extraction, stacked denoising auto-encoders (SDAE)
neural network is adopted to extract the fault features from nonlinear frequency spectrum. In this novel
method, four orders nonlinear frequency spectrum of each state of Permanent Magnet Synchronous
Motor (PMSM) are obtained by identification algorithm; Then, choosing suitable sampling points from
four orders frequency spectrum to construct high-dimensional data; Finally, stacked denoising auto-
encoders (SDAE) neural network is designed to realize the output of fault classification. Simulations indi-
cate that the proposed method has good real-time performance and high diagnosis accuracy.
Ó 2019 Elsevier Ltd. All rights reserved.
1. Introduction
Complex electromechanical system is an important support of
manufacturing industry and has been widely used in aerospace,
transportation, energy and other fields. The system contains many
non-linear units, which are interrelated with each other and have
strong uncertainty and non-linear characteristics. In addition, such
equipment often runs with heavy loads in harsh environments,
which is prone to break down. Therefore, it is important to diag-
nose the fault of complex nonlinear systems.
Up to now, many scholars have done a lot of researches on
equipment fault diagnosis, and their research methods can be
divided into three categories: the analytical model-based
approach, the signal-based approach and knowledge-based
approach. In terms of fault diagnosis research based on model
analysis, Ref. [1] established the fault model of mobile robot actu-
ator, and carried out fault detection by calculating the error
between the measured value and estimated value of the model;
Ref. [2] established a 10-DOF coupling non-linear dynamic model
of gear shaft-bearing transmission system, then studied the influ-
ence of contact temperature, fractal backlash and random load
on the dynamic characteristics of gear transmission system. In
terms of fault diagnosis research based on signal processing. In
terms of fault diagnosis research based on knowledge, Ref. [3] pro-
posed an optimal expert diagnosis system for the transient detec-
tion of commercial cage induction motor bar broken fault. Ref. [4]
proposed a circuit fault prediction method based on Grey Theory
and expert system to detect circuit fault; Ref. [5] proposed a
rule-based knowledge system, which used expert knowledge and
machine learning to diagnose the fault of CPS system. Among the
three fault diagnosis methods mentioned above, the biggest prob-
lem of model analysis method is that the accurate model of object
is needed, however, it is difficult to establish model for complex
system. Although there is no need to establish model for signal-
based approach and knowledge approach, the two methods need
to extract signal features and build knowledge base respectively,
which is not comprehensive for a dynamic system.
In recent years, the fault diagnosis method of deep learning has
attracted experts’ attention Ref. [6] proposed a new multi-scale
convolution neural network structure, which integrated
multi-scale learning into the traditional convolution neural net-
work system to realize multi-scale extraction and classification of
fault signals of wind turbine gearbox; Ref. [7] introduced batch
normalized optimization method into deep neural network to real-
ize fault diagnosis of bearing and gearbox; Ref. [8] introduced
adversarial learning into convolutional neural network, and
proposed a new deep adversarial convolutional neural network
for intelligent monitoring and fault diagnosis of industrial equip-
ment. Other scholars used other deep learning methods, such as
https://doi.org/10.1016/j.measurement.2019.107190
0263-2241/Ó 2019 Elsevier Ltd. All rights reserved.
⇑
Corresponding author.
E-mail address: chenlerui5566@stu.xjtu.edu.cn (L. Chen).
Measurement 151 (2020) 107190
Contents lists available at ScienceDirect
Measurement
journal homepage: www.elsevier.com/locate/measurement