Decentralized diagnostics based on a distributed micro-genetic algorithm fortransducer networks monitoring large experimental systemsP. Arpaia, P. Cimmino, M. Girone, G. La Commara, D. Maisto, C. Manna, and M. Pezzetti Citation: Review of Scientific Instruments 85, 095103 (2014); doi: 10.1063/1.4894210 View online: http://dx.doi.org/10.1063/1.4894210 View Table of Contents: http://scitation.aip.org/content/aip/journal/rsi/85/9?ver=pdfcov Published by the AIP Publishing Articles you may be interested in Use of fuzzy inference system for condition monitoring of induction motor AIP Conf. Proc. 1482, 441 (2012); 10.1063/1.4757510 Quantitative analysis of terahertz spectra for illicit drugs using adaptive-range micro-genetic algorithm J. Appl. Phys. 110, 044902 (2011); 10.1063/1.3624737 Genetic Algorithm Based Simulated Annealing Method for Solving Unit Commitment Problem in Utility System AIP Conf. Proc. 1298, 677 (2010); 10.1063/1.3516402 Computational Methods for Decentralized TwoLevel 0–1 Programming Problems through Distributed GeneticAlgorithms AIP Conf. Proc. 1285, 1 (2010); 10.1063/1.3510547 PRACTICAL GENETIC ALGORITHM BASED OPTIMAL CAPACITOR PLACEMENT FOR LOSS REDUCTIONAND VOLTAGE REQULATION IN DISTRIBUTION SYSTEMS AIP Conf. Proc. 1239, 64 (2010); 10.1063/1.3459788

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitationnew.aip.org/termsconditions. Downloaded to IP:

140.164.14.152 On: Wed, 15 Oct 2014 15:37:09

REVIEW OF SCIENTIFIC INSTRUMENTS 85, 095103 (2014)

Decentralized diagnostics based on a distributed micro-genetic algorithmfor transducer networks monitoring large experimental systems

P. Arpaia,1,2 P. Cimmino,1,2 M. Girone,1,2 G. La Commara,1 D. Maisto,3 C. Manna,4

and M. Pezzetti21Department of Engineering, University of Sannio, Benevento 82100, Italy2European Organization for Nuclear Research (CERN), Department of Technology, Geneva, Switzerland3Institute of High-Performance Computing and Networking, National Research Council, 80131 Naples, Italy4Cork Constraint Computation Centre, University College Cork, Cork City, Ireland

(Received 15 April 2014; accepted 18 August 2014; published online 8 September 2014)

Evolutionary approach to centralized multiple-faults diagnostics is extended to distributed transducernetworks monitoring large experimental systems. Given a set of anomalies detected by the trans-ducers, each instance of the multiple-fault problem is formulated as several parallel communicatingsub-tasks running on different transducers, and thus solved one-by-one on spatially separated par-allel processes. A micro-genetic algorithm merges evaluation time efficiency, arising from a small-size population distributed on parallel-synchronized processors, with the effectiveness of centralizedevolutionary techniques due to optimal mix of exploitation and exploration. In this way, holisticview and effectiveness advantages of evolutionary global diagnostics are combined with reliabilityand efficiency benefits of distributed parallel architectures. The proposed approach was validatedboth (i) by simulation at CERN, on a case study of a cold box for enhancing the cryogeny di-agnostics of the Large Hadron Collider, and (ii) by experiments, under the framework of the in-dustrial research project MONDIEVOB (Building Remote Monitoring and Evolutionary Diagnos-tics), co-funded by EU and the company Del Bo srl, Napoli, Italy. © 2014 AIP Publishing LLC.[http://dx.doi.org/10.1063/1.4894210]

I. INTRODUCTION

In monitoring large experimental systems, such as par-ticle accelerators, gravitational wave detectors, optical andradio telescopes, or nuclear fusion facilities, a large numberof distributed sensing and processing nodes is employed.1, 2

An effective real-time system, capable of handling massivedata, detecting anomalous deviations from nominal behav-ior, and diagnosing the corresponding causes promptly, isrequired. Resource distribution, communications limitations,poor scaling to configuration changes, and possible node/linkfaults create several design challenges to centralized diag-nostic paradigms.3 Alternative approaches can be classifiedin decentralized or distributed, according to the correspond-ing level of information sharing among the processors.4, 5 Indecentralized solutions, significant knowledge about the sys-tem to be diagnosed is shared among the processors. In dis-tributed solutions, processors host well-separated processes,by fitting better the requirements of complex systems.1–3

In literature, some distributed solutions aim at increasingthe computational efficiency,4, 6, 7 while others are mainlyfocused on working locally,4, 5, 8 by involving other sub-systems only when additionally information or hardwareresources are needed. However, all these approaches intro-duce ad hoc diagnostic architectures, devoted to specificproblems. Thus, a generic and application-independent ap-proach, as the Multiple Faults Diagnostic (MFD) problem inRef. 9, seems to be still missing for a distributed system. MFDhas been approached successfully as a combinatorial opti-mization problem, whose solution is a set of faults best ex-

plaining the detected anomalies. Nevertheless, for intercon-nected large experimental systems, fault identification is anon-polynomial-complexity combinatorial problem. Nature-inspired meta-heuristic algorithms (e.g., Genetic Algorithms,Particle Swarm Computation, Ant Colony Optimization, andso on)10, 11 proved to have capabilities of both overcoming lo-cal optima through parallel exploration of the solution spaceand driving the search towards the most promising solu-tions. However, their need for powerful computing resourcesseverely restricts their applications in distributed transducersnetworks. By considering these assumptions, proper meta-heuristic algorithms with very-low computational load andsatisfying optimization performance are Micro-Genetic Algo-rithms (MGA),12 a variant of Genetic Algorithms for smallpopulation sizes. They substantially reduce the number of to-tal evaluations for achieving the optimal solution and preventworking memory from overloads. Applications in experimen-tal physics are not rare, especially in nonlinear optics.13, 14 Inparticular, in Ref. 13, a MGA allows the optimum frequency,maximizing atomic effect stabilization by high-intensity laserfields, to be found. The fitness is computed by solving atime-dependent Schrodinger equation at different laser fieldintensities, on parallel processors communicating throughMessage Passing Interface (MPI), a widely used standardprotocol.15

In this paper, a distributed diagnostic procedure based ona micro-genetic algorithm for transducer networks monitoringcomplex physics systems is proposed. The well-settled evolu-tionary approach of centralized multiple-faults diagnostics is

0034-6748/2014/85(9)/095103/11/$30.00 © 2014 AIP Publishing LLC85, 095103-1

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitationnew.aip.org/termsconditions. Downloaded to IP:

140.164.14.152 On: Wed, 15 Oct 2014 15:37:09

095103-2 Arpaia et al. Rev. Sci. Instrum. 85, 095103 (2014)

extended to distributed transducer networks monitoring largeexperimental systems. In particular, in Sec. II, the problem ofthe multiple-fault diagnosis is formalized. In Sec. III, the dis-tributed micro-genetic algorithm for multiple-fault diagnos-tics is presented. Finally in Sec. IV, efficiency and accuracyperformance of the procedure is highlighted by two case stud-ies related to diagnosing (i) the cold box for the cryogenic sys-tem of the Large Hadron Collider at CERN and (ii) the mainedifice’s systems (e.g., anti-theft/anti-intrusion, air condition-ing, and so on) in remote building automation.

II. PROBLEM STATEMENT: MULTIPLE FAULTDIAGNOSIS

In monitoring complex experimental systems, the mostchallenging problem of finding the causes of several si-multaneous anomalies (Multiple-Fault Diagnosis)1 is statedas an abductive problem:2 the hypothesis, i.e., the set ofcauses, best explaining the observed anomalies, is to be found.Combinatorially, it can be stated as:2 “given a graph withN nodes numbered from 1 to N, an accurate permutationof N elements among 2N – in the worst case – verifyinga specified rule has to be found.” This problem is classi-fied as a hard combinatorial optimization, here formalizedas:

� the 4-tuple 〈D, M, C, M+〉, where D is a finitenonempty set of faults d, M is a finite set of anoma-lies (symptoms) m, C is a relation defined as a sub-set of D × M, pairing faults with the correspondinganomalies, and M+ = {m1, m2,. . . mu} is a subset ofM identifying the observed anomalies. Namely, (d,m)∈ C means that the fault d may cause the anomaly m.

� The diagnosis DI is the solution of the problem, de-fined as the subset of D identifying the faults eventu-ally responsible for the anomalies in M+.

� An a priori probability pj is associated to each fault djin the set of faults D. Values are assumed to exist andfaults in D are assumed to be statistically independent.

� Moreover, the relation C, pairing faults with the cor-responding anomalies, is assumed to be a matrix ofcausal strength cij, (such that 0 < cij < 1), representinghow frequently a fault dj causes the anomaly mi. For-mally, the causal strength cij is expressed as the con-ditional probability P (dj causes mi|dj), i.e., the fault djcauses the anomaly mi.

Symbolic, causal, and numeric probabilistic knowledgeis exploited to generate and assess plausible hypotheses aboutDI. At this aim, a relationship for calculating the “relativelikelihood,” denoted as L(DI, M+), of a diagnosis DI, giventhe observable anomalies M+, can be derived. The likelihoodis the product of three factors:

L(DI,M+) = L1L2L3

=⎧⎨⎩

∏m

i∈M+

⎡⎣1 −

∏d

j∈DI

(1 − cij )

⎤⎦

⎫⎬⎭

×⎧⎨⎩

∏d

j∈DI

⎡⎣ ∏

ml∈effects(d

j)−M+

(1 − cij )

⎤⎦

⎫⎬⎭

×⎧⎨⎩

∏d

j∈DI

[pj

(1 − pj )

]⎫⎬⎭ , (1)

where

� L1 is the likelihood that faults in DI cause the anoma-lies in M+. For a diagnosis not covering M+, L1 eval-uates to 0, thus forcing L to 0.

� L2 is the likelihood that faults in DI do not causeanomalies outside of M+. Ideally, L2 values close to1 are preferred.

� L3 is the likelihood that a highly probable fault dj con-tributes significantly in the overall likelihood of a di-agnosis DI containing dj.

L in (1) has to be maximized in order to find the mostprobable causes (faults) determining the observed anomalies.

III. DISTRIBUTED MICRO-GENETIC ALGORITHMFOR MULTIPLE-FAULT DIAGNOSTICS

In Sec. III A, the concept design, Sec. III B, the work-ing principle, and Sec. III C, the procedure of the distributedmicro-genetic algorithm for multiple-fault diagnostics areillustrated.

A. Concept design

The architecture of the proposed distributed systemfor automatic monitoring and multiple-fault diagnostics oflarge experimental systems is depicted in Fig. 1. Two dis-tributed functions are integrated at physical level on a net-work of smart transducers: (i) monitoring and fault detection

FIG. 1. Architecture of the distributed monitoring and multiple-faultdiagnostics.

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitationnew.aip.org/termsconditions. Downloaded to IP:

140.164.14.152 On: Wed, 15 Oct 2014 15:37:09

095103-3 Arpaia et al. Rev. Sci. Instrum. 85, 095103 (2014)

(continuous line in Fig. 1), for measuring and processing themain parameters of the system, as well as detecting possibleanomalies, and (ii) diagnostics (dashed line), for finding thecorresponding faulty causes. When one or more transducers(Monitoring Units in Fig. 1) detect a set of anomalies M+,the monitoring system expands its capabilities and a diagnos-tic process is triggered.

Each smart transducer acts as an isolated processor (Di-agnostic Unit in Fig. 1) executing a local independent task inorder to determine the final diagnosis DI. This multiple-faultdiagnostics function is based on a distributed Micro-GeneticAlgorithm (dMGA). Different algorithm instances, allocatedon the transducers of the monitoring network, evolve inde-pendently in parallel.16 At each T generations (immigrationinterval), the adjacent processes exchange their best indi-viduals synchronously, according to the elitist stepping-stonemigration model with a bi-directional ring topology.17 Thistopology is based on a classical coarse-grained approach:16

each instance on the Diagnostic Units is connected locally toother two instances, in order to exchange best individuals dur-ing the migration phase (bidirectional gray arrows in Fig. 1).For distributed networks monitoring large systems, this topol-ogy has several advantages: (i) improved efficiency owing tothe parallel process, shortening time execution, and fosteringpopulation diversity; (ii) broadcast communication networkemulated via point-to-point connections; (iii) extension up togeographical distances; (iv) increased reliability and easy net-work re-modulation in case of a faulty node, because only thetransducers in other subnets sending messages to the faultynode are affected; and (v) very-low cost of implementationand maintenance.

dMGA communications exploit the message passing in-terface (MPI) concept,15 widely adopted to develop portableparallel programming. Most important features are: essentialvirtual topology, synchronization and communication func-tion among a set of processes mapping network nodes ina language-independent way, with language-specific syntax(bindings), plus a few language-specific features.

In the concept design of the dMGA, the difficulty of in-tegrating the diagnostic knowledge into a unique representa-tion for a complex, dynamic, and distributed system is solvedby twofold distribution mechanisms:18 (i) spatial, where theknowledge is integrated from different local diagnostic pro-cesses placed in different subsystems and (ii) semantic, wherethe knowledge is integrated from different fields of expertise,related to the system physics, structure, and so on.

In the architecture of Fig. 1, the diagnostic process is de-centralized (non-distributed) because each spatially delocal-ized processor shares the same information about the rela-tionships anomaly-faults. In particular, a processor not onlyhas detailed knowledge about its monitored sub-system butalso an abstract view of the neighboring subsystems and ofthe system as a whole. Cooperating processors diagnose faultsaffecting more than one subsystem. A node triggers the co-operation process locally, when it realizes that the anomaliescannot be explained only within its subsystem. The cooper-ation process is driven by a small amount of topological in-formation. The Remote Supervision Station (RSS in Fig. 1) isnot involved in such a diagnostic process and acts mainly as

a final user interface. This decreases diagnostic response timedramatically for large experimental systems. Conversely, fora centralized diagnostics, the size of the system description islinear in the processor number and execution time will usuallybe even worse than linear.19 Moreover, all observations haveto be transmitted to the central diagnosis machine, causing alarge communication overhead.

The diagnostic process so conceived turns out to beglobal, because the nodes exchange continuously informationabout all the anomalies and the same information about therelationships anomaly-faults. Conversely, the dMGA comput-ing for the solution search is distributed, because each pro-cessor scans the solution space independently from the others.The unique information shared among evolutions allocated ondifferent processors is the best solution.

B. Working principle

The working principle of the dMGA is highlighted byreferring to its main design issues: (i) the initialization,(ii) the knowledge coding, (iii) the fitness, and (iv) the op-erators.

1. Initialization

At the beginning, for each dMGA instance situated on thenode of the transducer network, each population of N individ-uals is sampled randomly through a pseudo-random genera-tor with different seeds. Consequently, each instance starts itssearch from a different place of the solution space.

2. Knowledge coding

Given a set of relieved anomalies M+ = {m1, m2, . . . ,mu} as input, the dMGA returns the solution diagnosis as asequence of multiple faults DIbest = {d1, d2, . . . , dv}. The in-dividuals are designed as binary strings with length v equalto the cardinality of DI. At each generation, evolutionary op-erators update the individuals by modifying their own genes(bits).

Each gene corresponds to one and only one fault in DI,thus each individual represents a potential multiple-fault diag-nosis causing the anomalies M+. The gene is expressed, i.e.,the bit value is “1,” if the corresponding fault is present in thediagnosis as a probable cause in the solution. Conversely, it is“0.” Each individual, corresponding to a diagnosis DI, is as-sessed through the fitness for checking its attitude to representthe sequence of faults causing M+.

3. Fitness

Each individual, representing a potential diagnosis DI, isassessed by means of (1) for estimating the correspondencebetween the diagnosis it encodes and the most likely sequenceof faults causing M+ found until now. The likelihood L(DI,M+) is assessed by exploiting the matrix C reporting the u × vcausal strengths anomalies/faults. During the evolution, an in-dividual is selected as the best until the diagnosis it encodeshas the greatest likelihood.

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitationnew.aip.org/termsconditions. Downloaded to IP:

140.164.14.152 On: Wed, 15 Oct 2014 15:37:09

095103-4 Arpaia et al. Rev. Sci. Instrum. 85, 095103 (2014)

4. Operators

On each node of the transducer network, the dMGA in-stances initialize their own populations composed of N indi-viduals and perform the optimization process by applying, ateach generation, for a maximum number Gmax, the followingsix evolutionary operators:

1. Elitism: Memory about the best solution achieved so faris kept by introducing the survival of the best individ-ual. Elitism reserves a place in the offspring for the in-dividual of the current generation with highest fitness,by avoiding that other evolutionary operators affect itsgenome.20

2. Tournament selection: Individuals are selected for par-ticipating to a mating pool in order to generate the off-spring for the next generation. The selection relies onseveral tournaments among groups composed of a num-ber of individuals (the tournament size parameter, equalto N/2 in dMGA) randomly selected without replace-ment from the population. In each tournament, the indi-vidual with higher fitness wins and becomes an elementof the parent population for reproduction. Generally, theneeded tournaments number is equal to the populationsize, but in dMGA it is equal to N − 1 owing to theelitism.

3. One-point crossover: It is a very common crossovervariant for GAs. Initially, a unique genome point israndomly chosen with a probability PC for two parentindividuals. Subsequently, the part of the individuals’genome situated after the point is swapped between theparent individuals, by reproducing two child individuals.

4. Bit-flip mutation: Usually, mutation is not used in MGAbecause the diversification of the population relies onre-initialization operator. However, in dMGA, it resultsvery useful to increase population and thus the so-lution diversity. At this aim, the genome bits of thechild individual are flipped according to a probabilityPM. In practice, a bit is probabilistically inverted in itscomplementary.

5. Re-initialization: This is a peculiar operator of MGAs.12

Its dMGA application supports the exploration phase,otherwise penalized by the small size of the popula-tion. At each generation, the re-initialization operatorchecks if the genome of the highest-fitness individualhas a bit number less than a given percentage Hmicro fromthe other population elements. In this case, a nominalconvergence is relieved and the operator applies a newinitialization. Otherwise, the re-initialization operator isapplied when evolutionary process reaches a number ofinner generations (or micro-cycle) equal to Gmicro.

6. Migration: Solutions achieved in different evolutionaryalgorithms running in parallel can be exchanged bymeans of the migration operator. This permits singleexecutions to share their search spaces by generating,as a consequence, a global search process involving thedistributed algorithm as a whole. The dMGA migrationoperator is based on the abovementioned classical elitiststepping-stone migration.17 Specifically, at every fixedgeneration interval T, the copies of the best local indi-

viduals migrate between neighboring node processorsDU. Once reached a neighboring processor, the copyreplaces an individual randomly selected among theindividuals of the host population but differing from thelocal best one.

C. Procedure

dMGA consists of a set of MGAs instances assigned todifferent smart transducers and provided with the above oper-ators. These instances run in parallel in a folded bi-directionalring topology with a process casually set as master. The mas-ter process acts as an interface to the user, e.g., the RemoteSupervision Station (RSS of Fig. 1): it collects the current lo-cal best solutions of the “slave” processes and saves the bestelement at each generation. Moreover, it compares this lattersolution to the best found so far, saves the best among them,and transmits them to RSS.

The dMGA procedure includes the following steps(Fig. 2):

1. For each MGA instance allocated on the DU of a smarttransducer, a small-size population is initialized.

2. For each individual of the instance, the fitness value isassessed.

3. For each instance, if the nominal convergence subsistsor a micro-cycle halts, the local best individual is clonedvia the elitism operator and the rest of the local popu-lation is re-initialized. Afterword, the procedure comesback to step 2.

4. In each local population of an instance, the best localindividual is cloned by elitism and added to the localoffspring.

5. Locally, for each instance, individuals are chosen bymeans of the selection operator for reproduction.

6. Both crossover and mutation operators are applied toeach group of reproductive individuals selected in thedifferent dMGA instances in order to generate localoffspring.

7. The fitness of the new individuals belonging to every lo-cal offspring is evaluated.

8. In each dMGA instance, the local best individual is sentto the neighboring instances.

9. Each dMGA instance receives the copies of the best in-dividuals sent by the neighboring instances.

10. In each local population, the received individuals replaceother ones randomly chosen but different from the localbest individual.

11. If a halting condition is satisfied (number max of genera-tions, reached convergence, and so on), the correspond-ing dMGA instance halts; otherwise, the procedure re-turns to step 3.

IV. PROOF DEMONSTRATION

In the following, two case studies of validation of thedMGA are reported, related to (i) a simulation analysis fordeveloping cryogenics advanced diagnostics for the LargeHadron Collider at CERN and (ii) an experimental applica-tion for diagnosing technical plants in edifices.

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitationnew.aip.org/termsconditions. Downloaded to IP:

140.164.14.152 On: Wed, 15 Oct 2014 15:37:09

095103-5 Arpaia et al. Rev. Sci. Instrum. 85, 095103 (2014)

FIG. 2. Procedure of dMGA.

A. Simulation case study: LHC cryogenicsdiagnostics developments

The dMGA was applied at CERN to a first case studyfor analyzing possible future developments in the cryogenicequipment for 1.8 K refrigeration units. The cooling capacitybelow 2.0 K for the superconducting magnets of the LargeHadron Collider is provided by 8 refrigeration units of 2.4 kWby IHI-Linde and Air Liquide.

This case study is devoted to the diagnostics of onerefrigeration unit by IHI-Linde composed by (Fig. 3):21

(i) a warm compression station (WCpS), including an oil-lubricated screw compressor (WCp), with the associated oilremoval system (ORS); (ii) a cold compressor box (CCB),

FIG. 3. Architecture of the 1.8 K refrigeration unit by IHI-Linde.

including mainly a train of cold compressors (CC1-4), 2 heatexchangers (Hx1-2), a phase separator (Ph. Sep.), and 2 turbo-expanders (Tu1-2); (iii) the interfaces with the LHC (headerB); and (iv) a 4.5 K refrigerator (headers C and D).

1. Diagnostics problem

In the current situation at CERN, the refrigeration unitis handled by a monitoring and control system. Several sen-sors are linked to the PLCs, used to control the devices andto send the measured values to the general CERN monitoringsystem. This simulation case study is aimed at assessing theperformance of a future advanced diagnostic function to be in-tegrated in the monitoring system to support the operators.22

Most critical devices in the refrigeration unit are: (i) inthe warm compression station, the oil removal system, and(ii) in the cold compressor box, the compressors CC1-4 andthe cryogenic turbines Tu1-2. The WCp is a screw compres-sor, using a particular oil (Breox) to increase the tightness ofthe entire process. Before sending the compressed helium tothe cold compressor box, all the possible traces of oil haveto be removed up to a residual of only few ppb. The coldcompressors installed in the box are critical because they arecomplex systems relying on active magnetic bearings for shaftlevitation. Particular attention must be paid also to the cryo-genic turbines, because they cannot correctly run in presenceof impurities in the helium flow (Breox, water or nitrogen).Furthermore, it is important to properly regulate the break sys-tem, aimed at dissipating the mechanical energy produced bythe expansion process.

The distributed diagnostic system (Fig. 4) was conceivedto carry out the diagnostic task locally. The achieved diagno-sis response is sent to the CERN monitoring system (actingas Remote Supervision Station), while the PLCs are still usedfor the control process.

2. dMGA configuration

For this case study, dMGA instances, logically connectedin a stepping-stone bi-directional ring topology, were realized.The dMGA was implemented in language C and was exe-cuted on 3 virtual machines (64 bit architecture, 2 cores, 2 GBRAM). The Linux library Message Passing Interface was usedfor passing messages by Secure Shell protocol between themachines.

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitationnew.aip.org/termsconditions. Downloaded to IP:

140.164.14.152 On: Wed, 15 Oct 2014 15:37:09

095103-6 Arpaia et al. Rev. Sci. Instrum. 85, 095103 (2014)

FIG. 4. Architecture of distributed monitoring and multiple-fault diagnosticsproposed for LHC cryogenics at CERN.

The following parameters configuration was used: num-ber of machines: 3; number of runs: 10; population size: 15;migration rate: 10; crossover probability: 0.8; mutation proba-bility: 0.05; maximum number of generations: 100; and max-imum number of micro-generations: 10. A centralized bruteforce algorithm was also implemented in MATLAB and exe-cuted on a pc with i7 processor (3.40 GHz) with 8 GB RAM.This algorithm consists of systematically enumerating all thecandidates for the solution and searching for the maximumlikelihood.

3. Proof-of-principle scenario

The anomalies and faults vectors (Tables I and II, re-spectively), compiled according to the plant engineer’s expe-rience and requirements, include 80 anomalies and 55 faults.The a priori probabilities pj and the causal strengths cij wereidentified from a statistical analysis based on engineers andoperators knowledge and historical faults. All the faults ofTable II are critical because they definitely lead to system

TABLE I. Anomalies vector.

Anomalies

Water flow low A1 Bearings currents high CC2 A41Air pressure low A2 No sensors signal CC2 A42Voltage low A3 Axial force high CC2 A43Vacuum pressure high A4 Unbalance level bearing A high CC2 A44Water contamination A5 Unbalance level bearing B high CC2 A45Nitrogen contamination A6 Motor currents high CC2 A46Breox contamination A7 Motor temperature high CC2 A47Temperature brake circuit Tu1 High A8 Motor voltages high CC2 A48Temperature brake circuit Tu2 high A9 Levitation lost CC3 A49Pressure brake circuit Tu1 high A10 Bearings currents high CC3 A50Pressure brake circuit Tu2 high A11 No sensors signal CC3 A51Pressure Tu1 bearings low A12 Axial force high CC3 A52Pressure Tu2 bearings low A13 Unbalance level bearing A high CC3 A53Temperature Tu1 bearings low A14 Unbalance level bearing B high CC3 A54Temperature Tu2 bearings low A15 Motor currents high CC3 A55Inlet pressure Tu1 low A16 Motor temperature high CC3 A56Inlet pressure Tu2 low A17 Motor voltages high CC3 A57Inlet T1 mass flow low A18 Levitation lost CC4 A58Overspeed Tu1 A19 Bearings currents high CC4 A59Overspeed Tu2 A20 No sensors signal CC4 A60Inlet temperature Tu1 low A21 Axial force high CC4 A61Inlet temperature Tu2 Low A22 Unbalance level bearing A high CC4 A62Out temperature Tu2 low A23 Unbalance level bearing B high CC4 A63Inlet cold box mass flow low A24 Motor currents high CC4 A64Pressure (low pressure side) high A25 Motor temperature high CC4 A65Pressure (low pressure side) low A26 Motor voltages high CC4 A66Pressure (high pressure side) high A27 Helium mass flow low A67Pressure (high pressure side) low A28 Pressure (WCS low pressure side) high A68Dewar helium level high A29 Pressure (WCS low pressure side) low A69Inlet pressure CC1 high A30 Pressure (WCS high pressure side) high A70Levitation lost CC1 A31 Helium temperature (WCS high pressure side) high A71Bearings currents high CC1 A32 Oil temperature high A72No sensors signal CC1 A33 Oil mass flow low A73Axial force high CC1 A34 Compressor vibrations level high A74Unbalance level bearing A high CC1 A35 Pump currents high A75Unbalance level bearing B high CC1 A36 �Pressure on filters high A76Motor currents high CC1 A37 Oil contamination on absorber system A77Motor temperature high CC1 A38 Motor currents high A78Motor voltages high CC1 A39 Bearings temperature high A79Levitation lost CC2 A40 Motor temperature high A80

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitationnew.aip.org/termsconditions. Downloaded to IP:

140.164.14.152 On: Wed, 15 Oct 2014 15:37:09

095103-7 Arpaia et al. Rev. Sci. Instrum. 85, 095103 (2014)

TABLE II. Faults vector.

Faults

Water cooling system fault F1 Compressor motor faults CC2 F29Valves air system fault F2 Controller faults CC3 F30Power supply system fault F3 Position sensor failure CC3 F31Vacuum pumps fault F4 Magnetic bearing coil fault CC3 F32Cold box leakage F5 Compressor shaft faults CC3 F33Connections leakage F6 Compressor motor faults CC3 F34Turbine 1 inlet valve fault F7 Controller faults CC4 F35Filter turbine 1 fault F8 Position sensor failure CC4 F36Filter turbine 2 fault F9 Magnetic bearing coil fault CC4 F37Cryostat fault F10 Compressor shaft faults CC4 F38Joule Thomson valves fault F11 Compressor motor faults CC4 F39Break system 1 fault F12 Bypass valve failure F40Break system 2 fault F13 Buffer system fault F41Helium pollution F14 Hand valve failure (low pressure side) F42Gas analyzer fault F15 Discharge valve failure F43Bearing pressure regulator 1 fault F16 Heat exchanger fault F44Bearing pressure regulator 2 fault F17 Hand valve failure (high pressure side) F45Dewar level sensor fault F18 Oil filter faults F46Dewar level circuit fault F19 Circuit leakage F47Controller faults CC1 F20 Oil level sensors faults F48Position sensor failure CC1 F21 Oil pump faults F49Magnetic bearing coil fault CC1 F22 Oil valves faults F50Compressor shaft faults CC1 F23 Absorbers’ filters faults F51Compressor motor faults CC1 F24 Compressor bearings faults F52Controller faults CC2 F25 Gearbox failures F53Position sensor failure CC2 F26 Compressor shaft faults F54Magnetic bearing coil fault CC2 F27 Motor faults F55Compressor shaft faults CC2 F28

failures (permanent interruptions). An interruption for a sin-gle refrigeration unit means a stop for the LHC as a whole.Moreover, also from a technical point of view, this case studyis a hard combinatorial optimization problem:2 55 faults giverise to 255 possible solutions of the MFD problem. For thisreason, it turns out to be intractable with a brute-force algo-rithm. To give an idea, after 10 days of continuous running,the algorithm was capable to assess only 7.3642 × 109 possi-ble solutions on a total of 3.6029 × 1016.

According to the literature about evolutionaryalgorithms,23 main performance indexes are effectiveness (ameasure of the quality solution within a given computationallimit) and efficiency (a measure of the amount of computingneeded to achieve a satisfactory solution). In this paper, theeffectiveness was calculated as the average likelihood of thebest in run solutions, while the efficiency as the average ofthe generation’s number corresponding to the best likelihoodsolution.

The dMGA was tested on 4 different cases, consisting ofdifferent scenarios up to 5 simultaneous faults: (i) two cases,related to the cold compressors (CC1-4) with magnetic bear-ings, owing to past reliability problems; and (ii) two relatedto the warm compression station and to the cold box.

4. Simulation results

The dMGA was executed to evaluate the best solution interms of the faults with their occurrence probability. In effec-

tiveness and efficiency tests, the dMGA worked according toconfiguration described in Sec. IV A 2. The results for the10 most significant scenarios are reported in Table III. Theanomalies and the faults are encoded as arrays of binary chro-mosomes, thus, in Table III, the array locations where the bit“1” is present are indicated. The solutions correspond to rea-sonable faults. As an example, in test case 2, if the low voltageanomaly (A3) is detected, the power supply system fault (F3)is diagnosed. The average likelihood (3.502 ×10−2) is quiteclose to the best one (3.890 × 10−2) and the average genera-tion’s number is reasonable (61).

The results of Table III highlight how the dMGA is ca-pable of reaching the best likelihood solution in a fairly goodnumber of generations and the average likelihood is close tothe best likelihood. As an example, in test case 4, the averagelikelihood is equal to the best likelihood: 3.189 × 10−4, andthe algorithm reaches the best solution in an average genera-tion’s number of 56, with a standard deviation of 17, over 10test repetitions.

B. Experimental case study: Diagnostics of buildings

The dMGA was validated experimentally in an applica-tion in the framework of the project MONDIEVOB.24 Sev-eral smart transducers are installed in an edifice for monitor-ing and diagnosing automatically the main building systems:anti-theft/anti-intrusion, air conditioning, electrical system,elevators and lifts, fire system, and so on.

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitationnew.aip.org/termsconditions. Downloaded to IP:

140.164.14.152 On: Wed, 15 Oct 2014 15:37:09

095103-8 Arpaia et al. Rev. Sci. Instrum. 85, 095103 (2014)

TABLE III. Anomalies and corresponding faults with related likelihood and indexes determined by the dMGA.

Test case Anomalies Faults Best likelihood Average likelihood Average generation number

1 A4 F4 3.890 × 10−2 (3.502 ± 1.167) 10−2 (61 ± 20)2 A3 F3 3.890 × 10−2 (3.516 ± 1.123) 10−2 (51 ± 15)3 A5, A6, A7, A12, A13, A19,

A20, A21, A22, A23, A30F10, F14 8.289 × 10−10 (7.460 ± 2.487) 10−10 (64 ± 21)

4 A4, A76, A77 F4, F51 3.189 × 10−4 (3.189 ± 0.000) 10−4 (56 ± 17)5 A1, A8, A9, A16, A17, A18,

A19, A21, A22, A23F1, F7, F11 9.026 × 10−10 (6.494 ± 3.873) 10−10 (56 ± 18)

6 A34, A35, A36, A49, A50, A51,A64, A65, A66

F23, F31, F39 5.202 × 10−8 (4.805 ± 1.192) 10−8 (64 ± 17)

7 A3, A4, A67, A68, A69, A70,A76, A77

F3, F4, F41, F51 4.139 × 10−9 (2.942 ± 1.829) 10−9 (64 ± 17)

8 A1, A8, A9, A16, A17, A18,A19, A21, A22, A23, A78, A80

F1, F7, F11, F55 1.131 × 10−11 (8.065 ± 4.960) 10−12 (61 ± 17)

9 A4, A64, A65, A66, A67, A68,A69, A70, A73, A74, A75,

A76, A77

F4, F39, F41, F49, F51 1.169 × 10−12 (8.313 ± 5.161) 10−13 (69 ± 16)

10 A3, A5, A6, A7, A12, A13, A19,A20, A21, A22, A23, A29, A30,

A67, A73

F3, F10, F14, F19, F47 3.300 × 10−15 (1.540 ± 1.226) 10−15 (62 ± 21)

In the dMGA validation, the following edifice sub-systems were monitored owing to their criticalities, by in-stalling a network of smart transducers with the correspond-ing diagnostic units (Fig. 5): (i) Lift and elevators (LDU),by measuring all the parameters of the related electrome-chanical plant, as well as what is needed for the compliancewith EU safety standards (e.g., UNI-EN 81.28) and directives;(ii) Heating plant (HDU), a classical primary-secondarypumping system, by monitoring the thermo-fluid-dynamicquantities such as water temperature and pressure in primary(generator to the storage tank) and secondary (storage tank tousers) heating circuits (HDU); (iii) Air handling (ADU), bymonitoring the parameters (air temperature in the main circuitand in the heat recovery circuit) of the unit for air treatment,as well as of its humidifier and heat recovery circuit; (iv) Elec-trical/Alarm (EDU), by monitoring all the quantities of elec-trical, fire, and anti-theft plants; and (v) (DDU) Domotic, bymeasuring room temperature, humidity, comfort parameters,and so on. Finally, the Remote Supervision Station (RSS) isa software-structured user interface, for providing a remoteview of all the monitored plants and the diagnostics outputfor data storage and maintenance.

FIG. 5. Architecture of distributed monitoring and multiple-fault diagnosticsfor buildings.

1. Diagnostics problem

For each of the above subsystems, main faults have to bediagnosed on the basis of the corresponding anomalies. Asan example (Table IV): (i) for the electrical system (A1-13,F1-4), a power electrical black out, or a fault to the line trans-former have to be diagnosed by starting from the detectionof problems on the voltage lines R,S,T vs the neutral N, or atoo low power factor cos(fi); (ii) for the heating plant (A14-21, F5-14), a failure to the hydraulic separator between pri-mary and secondary circuits is to be diagnosed when an over-heating is detected on the primary circuit in summer position;(iii) for the air treatment (A22-25, F15-24), a faulty humidis-tat or a drift in the water temperature sensor of the secondarycircuit have to be explained by starting by an apparent over-heating in the heat recovery air.

2. Experimental setup

Each smart transducer includes (Fig. 6): (i) analog anddigital output and interfaces for sensor’s data acquisition;(ii) ARM9/Unix OS-based single board computer platform(ARM 9, 200 MHz, 32 MB RAM); (iii) wireless peripheralfor local network interface; and (iii) modem interface for re-mote GPRS channel communication.

In Fig. 7, a particular of the smart transducer hosting theDomotic Diagnostic Unit, with a fan-coil actuator based on atransistor ULN2003A and a connector DB25 for temperatureand solar radiation sensors, is highlighted. The dMGA wasimplemented in language C.

The following parameters configuration was used: num-ber of machines: 3; number of runs: 10; population size: 24;migration rate: 10; crossover probability: 0.8; mutation prob-ability: 0.2; maximum number of generations: 50; and maxi-mum number of micro-generations: 5.

The centralized brute force algorithm was also in thiscase study implemented in MATLAB and executed on a pc.

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitationnew.aip.org/termsconditions. Downloaded to IP:

140.164.14.152 On: Wed, 15 Oct 2014 15:37:09

095103-9 Arpaia et al. Rev. Sci. Instrum. 85, 095103 (2014)

TABLE IV. Anomalies and faults vectors (RST: electrical phases, N: neutral, cos(fi): power factor, ATU: airtreatment unit).

Anomalies Faults

Current break phase RN A1 Electrical power black out F1Current break phase SN A2 Electrical overload F2Current break phase TN A3 Transformer fault F3High voltage phase RN A4 Loads with power factor correction F4High voltage phase SN A5 Pressure reducer fault F5High voltage phase TN A6 Water losses F6Low voltage phase RN A7 Air in heating circuit F7Low voltage phase SN A8 Heat pump fault F8Low voltage phase TN A9 Hydraulic separator failure F9Low cos(fi) phase RN A10 Fault valve secondary circuit F10Low cos(fi) phase SN A11 Circulation pump primary circuit failure F11Low cos(fi) phase TN A12 Fault pump secondary circuit F12High overbalance electrical loads A13 Dirty filters primary circuit F13High pressure primary circuit A14 Dirty filters secondary circuit F14Low pressure primary circuit A15 Fault thermoregulation ATU F15High pressure secondary circuit A16 Faulty valve ATU F16Low pressure secondary circuit A17 No water supply F17High temp. primary circ. – summer A18 Pump failure humidification F18Low temp. primary circ. – summer A19 Battery failure humidification F19High temp. primary circ. – winter A20 Faulty humidistat F20High temp. primary circ. – winter A21 Water temperature sensor fault primary circuit F21High temp. air fan ATU A22 Water temperature sensor fault secondary circuit F22Low temp. air fan ATU A23 Air temperature sensor fault recovery F23High temp. recovery air ATU A24 Supply air temperature sensor fault F24Low temp. recovery air ATU A25

FIG. 6. Smart transducer with embedded ARM9 TS7200, interface PC104auxiliary board, and DB25 connector for analog and digital I/O.

FIG. 7. Domotic Diagnostic Unit with fan-coil actuator based on transis-tor ULN2003A with DB25 connector for temperature and solar radiationsensors.

In Fig. 8, the finite state automaton for the monitoring unit oflifts and elevators plants is illustrated.

3. Proof-of-principle scenario

Also in this case, the matrices for the relationshipsanomalies/faults C and a priori probabilities M of the pos-sible faults P were based on engineering experience and re-quirements. In particular, most of the faults were focused onelectrical, thermal, and air treatment units.

In Table IV, the considered 24 faults and the 25 anoma-lies are reported. The matrix C is large and sparse, with di-mension D × M = 24 × 25, i.e., 224 possible solutions of

FIG. 8. Finite state automaton for the monitoring unit of electrical, fire, andanti-theft systems.

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitationnew.aip.org/termsconditions. Downloaded to IP:

140.164.14.152 On: Wed, 15 Oct 2014 15:37:09

095103-10 Arpaia et al. Rev. Sci. Instrum. 85, 095103 (2014)

TABLE V. Single and multi-fault test results.

Test case Anomalies vector Faults vector Likelihood

1 A1, A2, A3 Brute Force F1 1.770 × 10−2

dMGA F1 1.766 × 10−2

2 A10 Brute Force F4 2.600 × 10−3

dMGA F4 2.572 × 10−3

3 A4, A10 Brute Force F3 1.742 × 10−4

dMGA F3 1.742 × 10−4

4 A15 Brute Force F6 5.860 × 10−2

dMGA F6 5.863 × 10−2

5 A15, A17, A23, A25 Brute Force F6, F15 3.100 × 10−3

dMGA F6, F15 3.118 × 10−3

the diagnostic problem. An anomalous transition of a pa-rameter triggers the start of the dMGA algorithm. The mas-ter unit (DDU) is responsible for vector transition detection,triggering the dMGA, and fault results encoding and trans-mission to RSS.

4. Experimental results

For a complete validation, single and multi-fault testcases were selected on different diagnostics problems of var-ious dimensions.

As an example of single-fault diagnosis, in the test case3 on the building electrical system (Table V), the anomaliesof an overvoltage line R vs neutral N with a simultaneous toolow power factor cos(fi) (A4-A10) were detected.

The dMGA diagnosed as the solution with best likelihood(1.742 × 10−4), a fault on the line transformer (F3), confirmedalso by the brute force algorithm. As multi-fault example, inthe Air Handling Unit, a too low pressure level on both the

primary (A15) and the secondary circuits (A17), as well as atoo low temperature level both on the fan (A23) and recovery(A25) air were detected.

The dMGA algorithm diagnosed as the fault with bestlikelihood (3.118 × 10−3) the simultaneous presence of bothwater losses (F6) and thermoregulation (F15) faults in theAir Handling Unit. This diagnosis was also confirmed by thebrute force algorithm.

As a matter of fact, for each test case, the brute force al-gorithm assessed the best solution in terms of fault vector andlikelihood. The dMGA was run on (i) 1, 2, and 3 DUs in par-allel with 100, 50, and 30 generations number; (ii) with 20runs; and (iii) with population size 24, 15, 10, and 7. Table Vreports the best dGMA result for the 5 test cases. The success-ful comparison between brute force and the dMGA likelihoodvalues points out a satisfying validation.

In Table VI, for the above 5 test cases, the percentageof success on 20 runs is highlighted at varying the numberof the involved diagnostic units, the number of generations,and the population size. The success condition is considered

TABLE VI. Success percentage at varying number of units, generations, and population size (lack of conver-gence in bold).

Population size

Test case No. of units No. of generations 24 15 10 7

1 1 100 90 90 45 152 50 95 80 60 103 30 95 80 40 15

2 1 100 85 50 40 102 50 85 45 40 403 30 90 75 30 10

3 1 100 95 80 25 102 50 95 80 40 253 30 95 75 55 25

4 1 100 30 35 40 152 50 55 45 25 153 30 60 45 20 5

5 1 100 35 30 30 202 50 35 20 15 53 30 45 15 20 5

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitationnew.aip.org/termsconditions. Downloaded to IP:

140.164.14.152 On: Wed, 15 Oct 2014 15:37:09

095103-11 Arpaia et al. Rev. Sci. Instrum. 85, 095103 (2014)

as the occurrence of the first run where the maximum likeli-hood value provided by the brute force algorithm is achieved.Results show how a decrease in the unit number up to 30%(e.g., owing to failures of 2 smart transducers on 3) can betaken into account preventively in the design by selecting 100generations and a population size of 24.

In fact, owing to this dMGA configuration, diagnosticssuccess turns out to be robust to processors loss, becausedMGA oversizing counterbalances the loss. Obviously, in thisextreme case, the parallelism is spent completely in redun-dancy, and thus only in reliability, without saving comput-ing resources. Conversely, a decrease in the population sizeup to 30% degrades drastically the performance, whatever thenumber of units and generations.

V. CONCLUSIONS

A diagnostic procedure, based on a distributed micro-genetic algorithm for transducer networks monitoring com-plex systems, has been proposed. The well-settled evolution-ary approach of centralized multiple-faults diagnostics is ex-tended to distributed networks.

The approach was tested at first in simulation at CERNfor LHC cryogeny diagnostics developments and then on thefield for the main systems of a building. Experiments provedthe following main innovations of the distributed micro-genetic algorithm: (i) improved efficiency by parallel pro-cess, shorter time execution, and higher population diversity;(ii) increased reliability and easy network re-modulation incase of a faulty node; and (iii) higher effectiveness and ef-ficiency of the dMGA in comparison with brute force algo-rithm.

ACKNOWLEDGMENTS

This work was supported by CERN trough the agreementKE1776/TE with the University of Sannio, and by the EU, inthe framework of POR 3.17 ICT of Regione Campania (Italy),whose support authors are acknowledged gratefully. The au-thors thank F. Cennamo for his inspiring suggestions, Maur-izio Marvaso of M2 srl, Giovanni and Stefania Del Bo, as wellas the Mondievob Team as a whole, for their “distributed”contribution.

1K. Obara, K. Kasahara, S. Hori, and Y. Okada, “A densely distributed high-sensitivity seismograph network in Japan: Hi-net by National Research In-stitute for earth science and disaster prevention,” Rev. Sci. Instrum. 76,021301 (2005).

2R. T. Shelton, J. H. Kamperschroer, L. J. Lagin, J. R. Nelson, and D. W.O’Brien, “Target diagnostic control system implementation for the Na-tional Ignition Facility,” Rev. Sci. Instrum. 81, 10E101 (2010).

3Y. Pencole and M.-O. Cordier, “A formal framework for the decentraliseddiagnosis of large scale discrete event systems and its application totelecommunication networks,” Artific. Intell. 164, 121–170 (2005).

4A. Subbiah and D. M. Blough, “Distributed diagnosis in dynamic fault en-vironments,” IEEE Trans. Parallel Distributed Syst. 15(5), 453 (2004).

5J. Kurien, X. Koutsoukos, and F. Zhao, “Distributed diagnosis of networkedhybrid systems,” in Proceedings of the AAAI Spring Symposium on Infor-mation Refinement and Revision for Decision Making: Modeling for Diag-nostics, Prognostics, and Prediction, 2002.

6A. Sheth, C. Hartung, and R. Han, “A decentralized fault diagnosis systemfor wireless sensor networks,” in Proceedings of the IEEE InternationalConference on Mobile ad hoc and Sensor Systems Conference, 2005.

7G. Provan, “A model-based diagnosis framework for distributed system,”in Proceedings of the 13th International Workshop on Principles of Diag-nosis, Semmering, Austria, 2–4 May 2002, pp. 16–24.

8N. Roos, A. ten Teije, and C. Witteveen, “Multi-agent diagnosiswith semantically distributed knowledge,” in Proceedings of the Belgium-Netherlands Artificial Intelligence Conference (BNAIC), Nijmegen,Netherland, 2003.

9Y. Peng and J. A. Reggia, Abductive Inference Models for DiagnosticProblem-Solving (Springer-Verlag, New York, 1990).

10M. Dorigo and L. M. Gambardella, “Ant colony system: A cooperativelearning approach to the traveling salesman problem,” IEEE Trans. Evolut.Comput. 1, 53–66 (1997).

11J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Proceedingsof IEEE International Conference on Neural Networks IV, Perth, WA, 27November–01 December 1995, pp. 1942–1948.

12D. E. Goldberg, “Sizing populations for serial and parallel genetic algo-rithm,” in Proceedings of the 3rd International Conference on EvolutionaryComputation, Nagoya, Japan, 1996, pp. 119–124.

13D. Barash, A. E. Orel, and V. R. Vemuri, “A genetic search in frequencyspace for stabilizing atoms by high-intensity laser fields,” J. Comput. In-form. Technol. 8, 103–113 (2004).

14D. L. Carroll, “Genetic algorithms and optimizing chemical oxygen-iodinelasers,” Develop. Theor. Appl. Mech. 18, 411–424 (1996).

15W. Gropp, E. Lusk, and A. Skjellum, Using MPI, 2nd Edition: PortableParallel Programming with the Message Passing Interface, MIT Press Sci-entific And Engineering Computation Series (Cambridge, 1999), Vol. 1.

16E. Alba and M. Tomassini, “Parallelism and evolutionary algorithms,”IEEE Trans. Evolut. Comput. 6, 443–462 (2002).

17H. Mühlenbein, “Evolution in time and space – The parallel genetic algo-rithm,” Foundation Genetic Algorithm. 316–337 (1992).

18R. Debouk, S. Lafortune, and D. Teneketzis, “Coordinated decentralizedprotocols for failure diagnosis of discrete event systems,” Disc. Event Dyn.Syst.: Theor. Appl. 10, 33– 86 (2000).

19L. Console, C. Picardi, and D. Theseider Dupr, “A framework for decen-tralized qualitative model-based diagnosis,” in Proceedings of the 20th in-ternational joint conference on Artifical intelligence IJCAI, San Francisco,USA, 6–12 January 2007, pp. 286–291.

20M. Gen and R. Cheng, Genetic Algorithms and Engineering Optimization(John Wiley and Sons, 2000).

21S. Claudet, P. Gayet, B. Jager, F. Millet, P. Roussel, L. Tavian, and U. Wag-ner, “Specification of eight 2400W@1.8 K refrigeration units for the LHC,”in Proceedings of the 18th International Cryogenic Engineering Confer-ence, Mumbai, India, 21–25 February 2000.

22R. Isermann, Fault-Diagnosis Applications (Springer-Verlag, Berlin,2011).

23A. E. Eiben and J. E. Smith, Introduction to Evolutionary Computing(Springer-Verlag, Berlin, 2003).

24Project MONDIEVOB, “Buildings Remote Monitoring and EvolutionaryDiagnostics,” Granted by EU in the framework of POR 3.17 ICT of Re-gione Campania (Italy).

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitationnew.aip.org/termsconditions. Downloaded to IP:

140.164.14.152 On: Wed, 15 Oct 2014 15:37:09