Computational Fault Diagnosis Technique forAnalog Electronic Circuits using Markov Parameters
V. Prasannamoorthy
1
, N. Devarajan
2
Department of Electrical Engineering, Government College of TechnologyCoimbatore, India
1
prasanna_gct1995@yahoomail.com
2
profdevarajan@yahoo.com
Abstract
—
In this paper, we propose a novel technique forthe detection and isolation of faults in analog electronic circuitsby the Simulation-Before-Test approach. The Markovparameters corresponding to each faulty state of the circuitunder test are computed. Following this, owing to the uniquenessof this numerical value to each of those faulty configurationsconsidered, a classifier may be designed such that it will becapable of isolating the faults by taking advantage of thatuniqueness. However, in this current work, we have restrictedour analysis to the determination of the parameters alone andtheir variations from the value obtained for the fault-free circuitstate. In cases where redundancies in the Markov parametersappear, the product of the Eigen values of the system matrix inthat configuration are computed to resolve ambiguity. TheSallen-Key bandpass filter and the Leapfrog filter have beenchosen to demonstrate the suitability of this approach topractical systems.
I.
I
NTRODUCTION
Diagnosis of faults in analog circuits is an unavoidableexercise in any industry involving electronic systems. Severalresearches [1], [3], [4] have addressed the issue of faultdiagnosis of analog electronic circuits at the system board andchip level. The research areas in this domain [5] encompasscomputational complexity, automatic test pattern generation,and design for testing process. Analog fault diagnosis iscomplicated by poor mathematical models, componenttolerances, nonlinear behaviour of components, and limitedaccessibility to internal nodes of the circuit under test. Thetraditional approaches to fault detection and diagnosis involvethe limit checking of some variables or the application of redundant sensors. More advanced methods are data-drivenprocess monitoring methods [8], [9] most heavily used inchemicals and manufacturing industries. Principal componentanalysis and partial least squares are multivariate statisticalmethods that generalize the univariate control charts that beenapplied for decades. Fisher discriminant analysis andcanonical variate analysis have also been used for diagnosispurposes. Other methods rely on analytical redundancy [7],[10], [11], i.e. the comparison of the actual plant behaviour tothat expected on the basis of a mathematical model. Thesemodels take their origins from chemical process control,where the traditional material and energy balance calculationsevolved into systematic data reconciliation and the detectionof gross errors [12]. The latter approach includes methods thatare more deterministically framed such as parity relationsfrom input
–
output model [6] and observers [7], [13] andthose formulated on more statistical basis [14] and parameterestimation [15].When analytical models are not readilyavailable, a correctly trained neural network can be used as anon linear dynamic model of the system [11], [16],[17].Sometimes, further insight is required as to the explicitbehaviour of the model involved and it is here that fuzzy [18]-[20] and even neuro fuzzy methods [21]-[23] come into theirown in fault diagnosis applications. Other authors have usedevolutionary programming tools to design observers [11], [24]and neural networks [25]-[27]. While working on faultdiagnosis, the Artificial Intelligence community was initiallyfocused on the expert system or knowledge-based approaches[28], where heuristics are applied to explicitly associatesymptoms with fault hypothesis. The short comings of a pureexpert system approach led to the development of model-based approaches based on qualitative models in form of qualitative differential equations, signed diagraphs, qualitativefunctional and structural models, etc., [8], [11], [29]. Most of the jobs that use knowledge-based methods work with modelsof system in the presence of the faults. This implies the needto construct a different model to each possible fault. Most of the time, it is not possible to obtain a model of the system witha particular fault, because the system could be damaged bythat fault, or because that might be dangerous to provoke thefaults or because not all possible faults can be provoked.Model-based fault diagnosis can be defined as the
determination of a system’s
faults by comparing the availablesystem measurements with a priori information represented by
the system’s mathematical
model through the generation of residual quantities and their analysis [30]. A complete model-based fault detection and isolation system must include atleast two modules: The residual generator where the plantbehaviours checked. Residuals are quantities that measure theinconsistencies between the actual plant variables and themathematical model. They are ideally zero, but they becomenonzero if the actual system differs from the model; this maybe a caused by faults, disturbances, noise and modelling errors.For a dynamic system, the residual generator is dynamic aswell. It may be constructed by means of a number of differenttechniques. An adequate design of the residual generatorallows the fault to be isolated, and therefore, classification of the residual vector into a specific fault case. An importantperformance characteristic of the residual generator is the faultsensitivity of the residuals that is, the ability of the generator
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 8, No. 1, April 2010195http://sites.google.com/site/ijcsis/ISSN 1947-5500