Fault Diagnosis of Analog Circuit Using Spectrogram And LVQ Neural
Network
Penghua Li
1
, Shunxing Zhang
1
, Dechao Luo
2
, Hongping Luo
1
,
1. Automotive Electronics Engineering Research Center, College of Automation, Chongqing University of Posts and
Telecommunications, Chongqing, 400065, China
E-mail: lipenghua88@163.com
2. Key Laboratory of Vehicle Emission and Economizing Energy, National Institute of Automotive Engineering, Chongqing, 400039
Abstract: This paper addresses a rened fault feature problem of analog circuit using a feature extraction technique
based on auditory feature. The proposed approach applies short-time fourier transform (STFT) to obtain the time and
frequency features of the fault responses being indicated separately by the cross and vertical axes in a spectrogram,
which gives much more rened description of the fault behavior. To reduce the computational complexity derived from
the high-dimensional texture features embedded in the spectrogram, the fault spectrograms are further processed by local
binary patterns (LBP) operator for obtaining low-dimensional fault features. Completing the parameter settings of the
network, the LBP feature vectors are fed to the learning vector quantization (LVQ) neural network for fault classication.
The numerical experiments about an active high-pass lter are carried out to indicate our approach has an acceptable
diagnostic rate with high accuracy.
Key W ords: Analog Fault Diagnosis, Spectrogram, Learning Vector Quantization, Local Binary Patterns
1 INTR ODUCTION
The analog systems are among the most unreliable and
least testable systems because of the unknown deviations
from the actual component tolerances and the complex
nature of the nonlinear fault mechanism [1]. The poor
fault models, component tolerances, and nonlinearity is-
sues make the analysis particularly complicated.
Over the past decades, considerable efforts have been
made for analog fault diagnosis at the system, board,
and chip levels [2]. Generally speaking, two broad
approaches, simulation-after-test (SAT) and simulation-
before-test (SBT), are used to obtain a circuit diagnosis in
the analog system [3]. The SAT approach identies faults
by estimating the circuit parameters from the measured cir-
cuit responses. The SBT approach obtains the fault iso-
lation by comparing the circuit responses elicited by the
predened test stimuli in the fault dictionary with the those
induced by different fault conditions in the actual situation.
Comparing with the SAT approach, the SBT approach has
some advantages when the topology of the circuit under
test (CUT) is complex. In fact, it allows reducing the test
time [4].
Among the SBT approaches, many neural-network-based
methods for the analog fault diagnosis, using the combi-
nation of various analytical techniques as preprocessor, are
discussed in academe. The authors in [1] collected data of
actual circuit which was preprocessed by wavelet decom-
This work is jointly supported by the National Natural Science
Foundation (61403053), the Youth Science and Technology Innovation
Talents Project of Chongqing (cstc2013kjrc-qnrc40005,CSTC2013kjrc-
tdjs40010), the Science and Technology Project of Chongqing Municipal
Education Commission (KJ1400404).
position, normalization and PCA to generate optimal fea-
tures for training the neural network. It ensured a simple
architecture for the neural network and minimizes the size
of the training set required for its proper training. In [2],
the authors divided circuit into modules, which, in turn,
were divided into smaller submodules successively. The
outputs of each module were addressed using wavelet anal-
ysis, PCA and data normalization for the fault classica-
tion. Reference [5] preprocessed the circuit responses by
maximal class separability based kernel PCA. The optimal
feature can reduce the computational burden of the neural
networks drastically. In [6], the authors employed wavelet-
based fractal analysis to obtain the features of the response
signals. The kernel PCA was used to reduce the dimension-
ality of candidate features for obtaining the optimal feature
inputs of linear ridgelet networks.
In fact, the preprocessing techniques of the aforementioned
fault diagnosis methods adopt the wavelet transform as the
means of feature extraction. To extract the features of the
circuit impulse responses, the corresponding output signals
are decomposed by wavelet transform. The approximation
coefcients are selected as the main features of impulse re-
sponse, yet the details representing the high-frequency con-
tents are discarded. In addition, the multivariate statistical
technique, PCA, should be carefully selected to generate
optimal features in the actual situation. In other words, this
method may be invalidated in these cases: if the probability
distributions of the investigated responses are not satised
with the Gaussian distribution or the exponential distribu-
tion; if the relationships between the principal component
vectors for the responses are not linear or orthogonal; if
the investigated responses have not a higher signal to noise
2673
978-1-4799-7016-2/15/$31.00
c
2015 IEEE