IntelligentAlarmClassificationBasedonDSmTAlbenaTchamovaJeanDezertAbstractInthis paper the critical issue of alarms classi-cation and prioritization (in terms of degree of danger) isconsideredandrealizedonthe base of Proportional ConictRedistribution rule no.5, dened in Dezert-Smarandache Theoryof plausible and paradoxical reasoning. The results obtained showthestrongabilityof this ruletotakecareinacoherent andstablewayfortheevolutionof all possibledegreesof danger,relating to a set of a priori dened, out of the ordinary dangerousdirections. AcomparisonwithDempstersruleperformanceisalso provided. Dempsters rule shows weakness in resolving thecases examined. InEmergencycaseDempsters ruledoes notrespond to the level of conicts between sound sources, leadingthat waytoungroundeddecisions. Incaseof lowest dangerspriority (perturbed Warning mode), Dempsters rule could causea false alarm and can deect the attention from the existing realdangerous source by assigning a wrong steering direction to thesurveillance camera.KeywordsAlarm classication; DSmT; DST; data fusion.I. INTRODUCTIONThe alarms classication and prioritization is a verychallenginganddifcult task. Theencounteredoverowingamount of alarms could become a serious source of confusionespecially in dangerous cases, when one needs to take a properimmediateresponse. Theproblemisreallycritical, becausethe information available for performing alarms processing isuncertain,imprecise, evenconicting. Therearecases, whensome of the alarms generated could be incorrectly interpretedasfalse, increasingthechancetobeignored, incasewhenthey are really signicant and dangerous. That way the criticaldelay of the proper response could cause signicant damages.Alot of work was done during the years, because theimportance of this problem was recognized since the 1960s, inwide world cases of surveillance: in industry (powerplants, oilreneries), the clinical alarms in medicine, civilian and mili-tary monitoring. Nowadays surveillance (military and civilian)and environmental monitoring systems are characterized withasmartoperationalcontrol, basedontheintelligentanalysisand interpretation of alarms coming from a variety of sensorsinstalled in the observation area. Many approaches have beenadoptedandapplied, addressingtheproblemincommon. In[1] a generic neuro-expert systemarchitecture for trainingneural networks in alarmprocessing is developed, whichis satisfactory when the training set covers enough rangeof scenarios. Anexpert systemwithtemporal reasoningforalarm processing is proposed in [2]. Fault detection and alarmprocessing in a loop system using a fault detection system ispresentedin[3]. In[4]theauthorsconsideramethodology,basedonbotharticial neural networksandfuzzylogicforalarm identication. The tasks of alarm processing, fault diag-nosis and comprehensive validation of protection performanceare discussed and resolved in [5] using knowledge-basedsystems andmodel-basedreasoningapproach. In[6] alarmprioritization, using fuzzy logic is developed to prioritize thealarmsduringalarmoodswhichwouldeasetheburdenofoperators with meaningless or false alarms. In case of multiplesuspicious signals, generated from a number of sensors in theobservedarea, the problemof alarmclassicationrequiresthemost dangerousamongthemtobecorrectlyrecognized,inorder todecideproperlywherethevideocamerashouldbe oriented. Because of uncertainty and conicts encounteredin signals data, one needs to process, analyze and inter-pret correctlyintimelymannerall suspicioussoundsignalsseparatelyat particular sensorslevelsintheobservedarea.Such kind of conicts could weaken or even mistake thedecisionabout the degree of danger ina critical situation.That is why a strategy for an intelligent, scan by scan,combination/updating of sounds data generated by each sensoris neededinorder toprovidethesurveillancesystemwithameaningfuloutput. Therearevariouswellknownmethodsfor combining information, which could be applied. The mostuseduntil nowDempster-ShaferTheory(DST)[9]proposesa suitable mathematical model for uncertainty representation,but its weak point in applications relates to the normalizationfactor, whichyieldstonon-adequateresultswhensourcestocombine are highly conicting. To overcome such drawback,weapplytheProportional Conict RedistributionRuleno.5(PCR5), denedinDezert-SmarandacheTheory(DSmT) ofplausibleandparadoxical reasoning[7]. It proposesapow-erful andefcient wayfor combiningandutilizingall theavailable information, allowing the possibility for conicts andparadoxes between the elements of the frame of discernment.AcomparisonwithDSTperformancebasedonDempstersrule of combination1is alsoprovidedinorder toevaluatetheabilityof DSmTtoassureawarenessabout thealarmsclassicationandprioritizationincaseofsoundsourcedatadiscrepanciesandtoimprovedecision-makingprocessaboutthe degree of danger. In section II we recall basics of DST and1This rule is also called Dempster-Shafer rule, and denoted DS for short.Originally published as Tchamova A., Dezert J., Intelligent Alarm Prioritization based on DSmT, IEEE Intelligent Systems IS2012, Sofia, Bulgaria, Sept. 6-8, 2012, and reprinted with permission.Advances and Applications of DSmT for Information Fusion. Collected Works. Volume 4381Dempstersrule. BasicsofPCR5fusionruleareoutlinedinsection III. Section IV relates to the decision making supportused in order to decide which sound source is most dangerous.In section V, we present the problem of alarms classicationandexaminetwosolutions tosolveit byusingPCR5andDempsters rule. In section VI, the evaluation and comparativeanalysis of both solutions are provided on a given simulationscenario, that includes three sensors, generating three types ofsignals (warning, alarm and emergency). Concluding remarksare given in section VII.II. BASICS OF DSTDST[9] proposesasuitablemathematical model for un-certaintyrepresentationLet = {1, 2, . . . , n}beaframeof discernment of a problem under consideration containing ndistinctelementsi, i=1, . . . , n. Abasicbeliefassignment(bba, alsocalledabeliefmassfunction) m(.): 2[0, 1]is a mapping from the power set of (i.e. the set of subsetsof ), denoted2, to[0, 1], that must satisfythefollowingconditions: 1) m() = 0, i.e. the mass of empty set (impossibleevent)iszero;2)

X2 m(X)=1,i.e.themassofbeliefis normalizedtoone. m(X) represents the mass of beliefexactly committed toX. The vacuous bba characterizing fullignorance is dened by mv(.) : 2[0; 1] such thatmv(X) =0if X=, andmv() =1. Fromanybbam(.),thebelieffunctionBel(.)andtheplausibilityfunctionPl(.)aredenedas X 2: Bel(X)=

Y |Y X m(Y )and Pl(X) =

Y |XY = m(Y ). Bel(X) and Pl(X) areclassicallyseenaslower andupper boundsof anunknownprobabilityP(X)ofX. Dempster-Shafer(DS)ruleofcom-bination[9] isamathematical operation, denoted , whichcorresponds to the normalized conjunctive fusion rule. Basedon Shafers model of the frame, the combination of twoindependent and distinct sources of evidences characterized bytheirbbam1(.)andm2(.)andrelatedtothesameframeofdiscernment is dened by mDS() = 0, and X 2\{}bymDS(X) = [m1 m2](X) =m12(X)1 K12(1)wherem12(X)

X1,X22X1X2=Xm1(X1)m2(X2) (2)correspondstotheconjunctiveconsensusonXbetweenthetwosourcesof evidence. K12isthetotal degreeof conictbetween the two sources of evidence dened byK12m12() =

X1,X22X1X2=m1(X1)m2(X2) (3)DSruleis commutativeandassociative. Theweakpointof this ruleis its behavior whenK121becauseit cangenerate unexpected(at least verydisputable) results [11].When K12=m12() =1, the two sources are said tobeintotal conict andtheircombinationcannot beappliedsinceDSruleismathematicallynot denedbecauseof 0/0indeterminacy [9].III. BASICS OF PCR5 FUSION RULEThe idea behindthe Proportional Conict Redistributionruleno. 5(see[7], Vol. 3) istotransfer conictingmasses(total or partial) proportionally to non-empty sets involved inthemodel accordingtoall integrityconstraints. Thegeneralprinciple of PCR rules is then to: 1) calculate the conjunctiveconsensus between the sources of evidences; 2) calculatethe total or partial conicting masses; 3) redistribute theconicting mass (total or partial) proportionally on non-emptysets involved in the model according to all integrity constraints.Under Shafers model assumption of the frame, the PCR5combination rule for only two sources of information isdened as:mPCR5() = 0 and X 2\ {}mPCR5(X) = m12(X)+

Y 2\{X}XY =[m1(X)2m2(Y )m1(X) + m2(Y )+m2(X)2m1(Y )m2(X) + m1(Y )] (4)wherem12(X)correspondstotheconjunctiveconsensusonXbetweenthetwosourcesandwhereall denominatorsaredifferent fromzero. All setsinvolvedintheformulaareincanonical form. All denominators are different from zero. If adenominator is zero, that fraction is discarded. No matter howbigor small theconictingmass is, PCR5mathematicallydoesabetter redistributionof theconictingmassthanDSsince PCR5 goes backwards on the tracks of the conjunctiverule and redistributes the partial conicting masses only to thesets involved in the conict and proportionally to their massesput intheconict, consideringtheconjunctivenormal formof the partial conict. PCR5 is quasi-associative and preservesthe neutral impact of the vacuous belief assignment.IV. DECISION-MAKING SUPPORTInthiswork, weassumeShafersmodel andweusetheclassical Pignistic Transformation[7], [10] totake a deci-sion about the mode of danger. The pignistic probability(Pign.Proba), also called the betting probability (BetP) isdened for A 2byBetP(A) =

XD|X A||X| m(X) (5)where |X| denotes the cardinality ofX.V. ALARMS CLASSIFICATION APPROACHOur approach for alarms classication assumes all the local-ized sound sources to be subjects of attention and investigationfor being indication of dangerous situations. The specicattributes of input sounds, emitted by each source, are sensorslevel processed and evaluated in timely manner for theircontribution towards correct alarms classication (in term ofdegreeofdanger). Theinput soundsattributesgeneratedbyeach sensor, at each time moment (scan) concern the frequencyof intermittence, fintand sound signal duration, Tsig. Aparticular relationship between the specic values offintandAdvances and Applications of DSmT for Information Fusion. Collected Works. Volume 4382associatedcorrespondingdegreeofdangerisestablished, i.eto map input specic sensor level data into the frame ofdiscernments, concerningthelevel of abstractionDegreeofDanger= {Emergency, Alarm, Warning}. Thenthe processconsists intemporal sensors level soundsignals attributeupdating on the base of PCR5 fusion rule. Our motivation forattribute fusion is inspired from the necessity to ascertain thedegree of danger, associated with all localized sound sourcesseparately, inordertoquicklyfocusonthemost dangerousalarm information and to take immediate and correct feedbackactions to decide properly where the video camera should beoriented. The applied algorithm considers the following steps:Wedenetheframeofexpectedhypothesesaccordingtothe respective degree of danger associated with the attributessspecicvaluesasfollows: = {1=(E)mergency, 2=(A)larm, 3=(W)arning}. Thehypothesiswithahighestpriorityis Emergency, followingbyAlarmandthenWarn-ing. These hypotheses are exclusive andexhaustive, henceShafers model holds and we work on power-set: 2={, E, A, W, E A, E W, A W, E A W}. A rule-base is dened in order to establish the relationshipsbetweenthesounds attributes associatedwithall localizedsources and corresponding degrees of danger, in the form:Rule 1: if attributes-type 1 then EmergencyRule 2: if attributes-type 2 then AlarmRule 3: if attributes-type 3 then Warningwhere attributes types 1, 2 and 3 could be specic sounds at-tributes values, which are informative enough to be processedandevaluatedfor their contributiontowardscorrect alarmsclassication. Inthisrulebaseattributes-type1isasoundsattribute, whichistypical for degreeof danger Emergency,attributes-type 2is typical for Alarm, attributes-type 3forWarning. Inourcasethefrequencyofintermittencies(ifthesignal is intermittent) fint, associated with the localized soundsources is utilized. Thenthefollowingspecicrule-baseisusedasaninput interfacetomapthesoundsattributes(socalledobservations)obtainedfromall localizedsourcesintonon-Bayesian basic belief assignmentsmobs(.):Rule1:if fint 1Hzthenmobs(E)=0.9andmobs(E A) = 0.1.Rule 2: iffint 5Hzthenmobs(A) = 0.7,mobs(A E) =0.2 andmobs(A W) = 0.1.Rule 3: iffint 0Hzthenmobs(W) = 0.6 andmobs(W A E) = 0.4.If thevalueof thesoundattributereceivedis closetotheparticular sound signal parameter for Emergency, our bbais constructed in way that it will consider the hypothesisEmergency andalsothe reasonable inthis case compositeproposition (EA), representing a possible partial uncertainty.If thevalueobtainedisclosetotheparticular soundsignalparameter for Alarm, our bba is constructed in way that it willconsider the hypothesis Alarm itself and also the reasonable inthat case composite propositions AE and AW. AssigningahighermassofbelieftoA EthantoA Wistotakecareabout thepossibilityfor Emergencycase. If thevalueobtained is close to the particular sound signal parameter forFig. 1. Scenario.Warning, ourbbaisconstructedinwaythat it will considerthehypothesis WarningandalsothecompositepropositionE A W, representing the case of full ignorance, in ordertotakecareabout possibilityfor AlarmandespeciallyforEmergency case. All the belief masses not already assigned tosingletons (E, A or W) are assigned to the reasonable partialuncertainties reecting the possible noise perturbations in theobserved information. At the very rst time moment k =0 we start with apriori basicbelief assignment (history) set tobeavacuousbelief assignment mhist(E A W) = 1 , since there is noinformation about the rst detected degree of danger accordingto sound sources. Combination of currently received measurements bbamobs(.) (for eachof locatedsoundsources), basedontheinput interface mapping, with a historys bba, in order toobtain estimated bba relating to the current degree of dangerm(.) =[mhist mobs](.). PCR5andDSaretestedintheprocess of temporal datafusiontoupdatebbas associatedwith each sound emitter.Flagfor anespeciallyhighdegree of danger has tobetaken, when during the a priori dened scanning period,themaximumPignisticProbability[7]isassociatedwiththehypothesis Emergency.For securitypurpose, it isveryimportant tokeepupdatingsequentiallytheestimationonehasonthestateof thetruemodes of sound emitters, even if they are in the lowestpriority mode (i.e. in warning mode only) in order to preventunexpected alarms changes.VI. SIMULATION SCENARIO AND RESULTSInour simulationscenario(Fig. 1) aset of threesensorslocatedat different distancesfromthemicrophonearrayareinstalled in an observed area for protection purposes, togetherwith a video camera [8]. It is assumed, that sensors areassembled with alarmdevices, as follows: Sensor 1 withSonitron, Sensor 2 with E2S, and Sensor 3 with System Sensorcompanies alarmdevices. Incase of alarmevents (smoke,ame, intrusion, etc.) the alarm devices emit powerful soundsignalswithvariousdurationandfrequencyofintermittencedepending on the nature of the event. dangerous signal source.These sensors are used for the purpose of estimation thelevelofdanger/threatforeachplacewheretheyarelocated.Data,obtainedfromeachsourceareprocessedandanalyzedat particular sensors level independently, in consecutive timeAdvances and Applications of DSmT for Information Fusion. Collected Works. Volume 4383Fig. 2. Sonitron, E2S, System Sensor Sound Characteristics.moments, with regard to all possible degrees of danger:1= (E)mergency, 2= (A)larm, and3= (W)arning.Doing this one could nd the rst suspicious moment, whenTable 1 Sound signal parameters.Continuous Intermittent-I Intermittent-II(Warning) (Alarm) (Emergency)fint= 0Hz fint= 5Hz fint= 1HzTsig= 10s Tsig= 30s Tsig= 60sthe situation could become eventually dangerous.The sound signals representing Warning, Alarm and Emer-gency, emitted from alarm devices, produced by Sonitron, E2SandSystemSensorcompaniesusedinoursimulation(Table1) are shown on Fig. 2. The rst (left) column of Fig. 2 relatestoSonitron, thesecondcolumntoE2S, andthethird(right)columnrelates toSystemSensor devices. Therst rowofthisgurerepresentsthesignal 1for Warning, secondrowrepresents signal 2, for Alarm, and the last third row representssignal 3, for Emergency case. The Alarm signal is intermittentwith a frequency of intermittencefint= 5Hzand a durationTsig=30s, so called type I. The Emergency sound signal isintermittent with a frequency of intermittence fint= 1Hz anddurationTsig= 60s, so called type II. The Warning signal iscontinuous withfint= 0HzandTsig= 10s.Oursimulationscenarioconsidersatruedegreeofdangerassociated with the sound sources as follows: Emergency modefortherst soundemitter, Alarmmodeforthesecond, andWarning mode - for the third one. The three sources are pro-cessed in parallel and because of possible sound perturbationswe assume that possible randomchanges canbe observedover the scans for a given mode. We therefore introducesomeswitchesbetweenthethreemodes Emergency, AlarmandWarningtosimulatewhat canhappeninpractice(whatwecall groundtruthanddisplayedwithblackplotsonournext gures 3 and 4. According to this, three main cases areestimated: The most interesting for us it is the estimation of dangerlevel bysensor1, associatedwithEmergencymode. Inour simulation, the The GroundTruthassociatedwithSensor 1 considers that during scans 13 the observationsgenerated support the Emergency mode (the highest levelof danger). Fromscan 4 to scan 6 the observationsgeneratedsupport theWarningmode(thelowest levelof danger). Fromscan 7toscan 30the observationsgenerated support again Emergency mode. Such kind ofscenariois important inthe real worldcases becausesourcesdatacanbedeterioratedbynoiseperturbationsandthereforesomepossibleconictsarisebetweenob-servationsfromscantoscan.WeassumethataconictoccursinsoundsdatabetweenEmergencyandWarningmodes, because it couldweakenstronglythedecisiontaken. It could become a reason to ignore the signicanceof out of ordinary, dangerous situation. Thesecondinterestingcaseconcernstheestimationofprobabilities of modes, associated with the sound emitter2workinginAlarmmode. TheGroundTruthhasbeena little bit changed with respect to the ground truthsimulated for sensor 1. We assume that during scans13 the observations generated support correctly theAlarmmode. Fromscan 4toscan 8theobservationsgenerated support the Emergency mode because of noiseperturbations. Fromscan9toscan30theobservationsgenerated support again correctly the Alarm mode. The third interesting case concerns the estimation of theprobabilityof modes, associatedwiththethirdemitterworking in Warning mode. In our simulation of thiscase, we considers that during scans 12 the observationsgenerated support correctly the Warning mode. Fromscan 3 to scan 5 the observations generated supportthe Emergency mode because of some possible noiseperturbations. Fromscan6toscan30theobservationsgenerated support again correctly the Warning mode.As a result of processing and analyzing sounds data,obtainedfromthethreesources, processedinparallel, oneestablishes at each scan, for each source the Pignistic probabil-ities, associated with all the considered modes of danger. Thedecisions should be governed at the video camera level, takenperiodically, depending on: 1) specicities of the video camera(timeneededtosteer thevideocameratowardalocalizeddirection); 2)timedurationneededtoanalyzecorrectlyandreliablythesequentiallygatheredinformation. Wechooseasareasonablesamplingperiodfor cameradecisions Tdec=20sec, i.e. at every 10th scan, we should establish the decisionabout the most probable mode of danger, associated with eachsoundsource, that waytodeclaredirectionsforsteeringthevideocamera. For our scenario, thedecisivescans will be10th, 20th, and30th. In the next two subsections we analyzethe performances of PCR5 and DS to conclude on their ability(or inability) tocorrectlyidentifythealarmmodes for theprioritization purpose.A. PCR5 rule performance for danger level estimation.Figure 3 shows the values of Pignistic Probabilities ofeach mode (Emergency, Alarm, Warning) associated with threesound emitters (1st source in Emergency mode, (subplot on thetop), 2nd source in Alarm mode (subplot in the middle), and3rd source in Warning mode, (subplot in the bottom)) duringthe all 30 scans. Each source has been perturbed with noises inAdvances and Applications of DSmT for Information Fusion. Collected Works. Volume 43840 5 10 15 20 25 3000.51SOURCE 1 EMERGENCY mode:PCR5 Rule Performance Ground TruthEmergencyAlarmWarning0 5 10 15 20 25 3000.51SOURCE 2 ALARM mode:PCR5 Rule Performance Ground TruthEmergencyAlarmWarning0 5 10 15 20 25 3000.51Scan numberPignistic ProbabilitiesSOURCE 3 WARNING mode:PCR5 Rule Performance Ground TruthEmergencyAlarmWarningFig. 3. PCR5 rule Performance for danger level estimation.accordance with the simulated Ground Truth, associated withparticularsoundsource. Theseprobabilitiesareobtainedforeach source independently as a result of sequential data fusionof mobs(.) sequence using PCR5 combinational rule. For eachsource, we analyze the probabilities of its modes obtainedwithBetPcomputedfromPCR5ruleandthecorrespondingdecisions for steering the camera at scans no. 10, 20, and 30.Decision taken by PCR5 rule at scan10:For source 1, associated with Emergency mode (Fig. 3, top-subplot), Pign.ProbaestablishedbyPCR5at scan10areasfollows:BetP(E) = 1.0,BetP(A) = 0, andBetP(W) = 0.During the rst scans one hasBetP(E)