View
214
Download
0
Category
Preview:
Citation preview
ORIGINAL ARTICLE
Comments on “enhancement of an efficientliveness-enforcing supervisor for flexiblemanufacture systems”
Daniel Y. Chao
Received: 1 May 2011 /Accepted: 1 September 2011 /Published online: 23 September 2011# Springer-Verlag London Limited 2011
Abstract Region theory can synthesize maximally permissivesupervisors by solving a set of inequalities based on themarking/transition-separation instances (MTSIs). It is infeasibleto solve these inequalities for either a sizable net or a small netwith a sizable initial marking. Huang et al. [1] propose novelcrucial MTSIs to reduce the number of MTSIs. Experimentalresults show that the proposed control policy is the mostefficient algorithm among the closely related approaches. Oneexample shows that it not only reaches all live states but alsoemploys fewer control arcs than that by Li et al. Huang et al.offer no hints on why it employs fewer control arcs. Thispaper develops theory to explain the physics behind.
1 Introduction
utilized as much as possible. This leads to strong competitionfor resources among process. Consequently, deadlocks occurdue to processes holding resources mutually waiting for othersto release their resources. Deadlock prevention [1–16] addsmonitors to problematic siphons to prevent them frombecoming insufficiently marked.
Uzam and Zhou [2, 3] apply region analysis to a well-known S3PR (systems of simple sequential processes withresources) in Fig. 1. The benchmark reaches 26,750 states,where 21,581 are legal, i.e., either good or dangerous states
(i.e., boundary states to reach forbidden regions). There are5,299 elements in the set of marking/transition-separationinstances (MTSI) denoted by Ω. This implies that 5,299 LPPs(linear programming problems) have to be solved to find anoptimal liveness-enforcing supervisor with 21,581 reachablestates in the controlled system. However, |Ω| (cardinality ofΩ) grows exponentially with respect to the size of a plantmodel. It is clearly infeasible to solve |Ω| LPPs for either asizable net or a small net with a large initial marking.
They further propose an iterative approach [4]. At eachiteration, a first-met bad marking (FBM) is singled out fromthe reachability graph of a given Petri net model. Theobjective is to prevent this marking from being reached viaa place invariant of the Petri net. A well-establishedinvariant-based control method is used to derive a controlplace. This process is carried out until the net modelbecomes live. The proposed method is generally applicable,easy to use, effective, and straightforward although its off-line computation is of exponential complexity. Huang et al.apply the FBM method to a well-known S3PR to achieve anear maximumally permissive control policy (21,562states). The more FBM, the more monitors and controlarcs will be added. The benchmark above requires moremonitors and arcs than the one in [5], where severalmonitors using the FBM method are lumped into one.
Li et al. [5] propose a two-stage policy as described below.In the first stage, a monitor is added for each elementarysiphon based on the method in [6, 13]. Next, the controllednet is still with some dead marking, and they utilize thetheory of regions to control the deadlock problem. An MTSIis associated with each illegal or forbidden marking. It statesthe MTSIs play a guard role to prevent all possible markingsfrom being entered into the dead zone. However, quite a fewof the equations must be solved by the MTSI stage [5]. Theresulting model is maximally permissive.
D. Y. Chao (*)National Chengchi University,Taipei City, Taiwan, Republic of Chinae-mail: yaw@mis.nccu.edu.tw
Int J Adv Manuf Technol (2012) 60:625–629DOI 10.1007/s00170-011-3628-4
Keywords Petri nets . Flexible manufacturing system .
Deadlock prevention
In an FMS, a set of processes share a set of costly resources. Tobe competitive and save the cost, these resources must be
Huang et al. [1] propose novel crucial MTSIs (CMTSIs)to reduce the number of MTSIs and hence the computationburden. In addition, the number of control arcs is fewer.Experimental results show that the proposed control policyis the most efficient algorithm among the closely relatedapproaches. Furthermore, the resulting model remainsmaximally permissive with no redundant monitors. It isquite time-consuming to remove redundant monitors usingthe method in [10]. However, Huang et al. offer no hints onwhy it employs fewer control arcs.
Intuitively fewer MTSI corresponds to fewer forbiddenmarkings and hence, fewer control arcs are employed asmentioned earlier. However, both models (Huang et al. [1]and Li et al. [5]) employ the same number of monitors andare maximally permissive. This paper develops a newtheory to explain the physics behind.
To shorten the paper, we assume the reader is familiarwith basic terminologies of Petri nets in [6].
2 Formal derivation
We first give an intuition as to why fewer control arcs areemployed in [1] followed by an example and a theory toexplian it in more detail. In the sequel, we restrict markingto the set of operation places as in [4] using the FBMapproach.
2.1 Intuitive explanation
The FBM method adds a monitor for each FBM. There maybe a number of FBM for a siphon, and each one needs amonitor. If one lumps these monitors into a single one,fewer control places and arcs can be employed. However,the modified (called controller) region [V] ([V] plus V formsa P-invariant [17]) is larger (than that of marked operationplaces) and affects the uncontrolled model more; hence itreaches fewer states. The MTSI method considers all FBMwhile the CMTSI considers only a few equivalentlylumping several FBM into the critical one. This roughlyexplains why the CTMSI method employs fewer arcs.
For the benchmark in Fig. 1, the FBM method by Uzamet al. results in the controlled model in Table 1. If onelumps V 1
6 ;V26 ;V
36 ;V
46
� �resp: V 1
9 ;V29
� �; V 1
10;V210
� �; V11;f�
V12;V12gÞ into a single V6 (resp. V9, V11), the resultingmodel reaches fewer (21,363) [16] states than the FBM one(21,562) in [4], but with 11 monitors and 54 control arcsfewer than 19 monitors and 120 control arcs reported in [8].
2.2 Example from [1]
The second example from [1] is shown in Fig. 2. Using thetwo-stage policy by Li et al., add monitor (places VS2, VS3,and VS4) in the first stage for each elementary siphon. Thepartially controlled net is called (N2L1, M0).
p26
WP11
7
p23
p20
p24
p21
p25
WP12
p22
2
2
p8 p12
p11
t7
t8
t9
t10
t6
t5
t4
t3
t2
t1
p13p9
p10
p7
p6
p2
p1
p4
p 3
t14
t13
t12
t11p15t19
t20
p17
t18
p16
t17
p18
p19
t15
p5
p14
WP3
WP2
t16
WP1
3
11
Fig. 1 Benchmark 1
626 Int J Adv Manuf Technol (2012) 60:625–629
It is worth to mention that eight MTSIs are found. Itstates the MTSIs play a guard role to prevent all possiblemarkings from being entered into the dead zone. However,quite a few of the equations must be solved by the MTSIstage [5]. Huang et al. [1], however, only employs oneCMTSI to find the final controlled model, which is muchbetter than the above eight sets of MTSIs required by Li etal. More MTSI results in more refined controlled model andhence requires more control arcs. Theoretically, the CMTSIshould employ fewer monitors than the MTSI method. Butboth methods employ six monitors. The next subsectionwill uncover this mystery.
2.3 Theory
All siphons for the net in Fig. 3 (synthesized from controlcircuits c1 ¼ VS2 t11 VS3 t5½ � and c2 ¼ VS4 t11 VS3 t5½ � basedon the theory in [15] in the partial controlled model inFig. 3) cannot be emptied based on the theory in [17].Siphon S8 ¼ p7; p11; p12; p17; p19;VS2;VS3f g (see Table 1 in[17]) is synthesized from the core subnet obtained byadding TP-handle HTP1[t11 p17 t10 p19 t7 VS3] with resourceplace r=p19 upon c3. S8 is emptiable as explained below.S8½ � ¼ p3; p5; p6; p9; p10f g, M0(p19)=1, M0(VS3)=3, andM 0 ( V S 2 ) = 1 . maxM S½ �ð Þ ¼ M0 CSð Þ ¼ 3þ 1 ¼ 4 a sexplained below (CS ¼ VS2;VS3f g is the set of controlplaces in S). Firing t9 once grabs one token from each of p19and V3. The other two tokens in V3 go to p10 and p6,
Table 1 Controlled model by Uzam et al. [8] for the benchmark in Fig. 1
S V(M0) V �S
�VS [VS]
S1 V1 (2) t9, t15 t10, t16 p13, p19S10 V2 (2) t3, t11 t4, t13 p2, p3, p8S18 V3 (2) t7, t17 t8, t18 p11, p17S16 V4 (2) t8, t16 t9, t17 p12, p18S15 V5 (3) t8, t15 t10, t17 p12, p13, p18, p19S4 V 1
6 ð5Þ t3, t12, t15 t5, t13, t17 p3, p8, p9, p12, p13, p18, p19S4 V 2
6 ð5Þ t4, t8, t11, t15 t5, t9, t13, t17 p2, p3, p9, p12, p18, p19S4 V 3
6 ð5Þ t4, t9, t11, t15 t5, t10, t13, t17 p2, p3, p9, p13, p18, p19S4 V 4
6 ð5Þ t4, t11, t15 t5, t13, t17 p2, p3, p9, p18, p19S17 V7 (5) t1, t17 t3, t8, t19 p6, p7, p11, p16, p17S19 V8 (3) t7, t16 t8, t17 p11, p18S20 V 1
9 ð4Þ t7, t15 t9, t17 p11, p12, p18, p19S20 V 2
9 ð4Þ t7, t9, t15 t8, t10, t17 p11, p13, p18, p19S21 V 1
10ð6Þ t3, t7, t12, t15 t5, t8, t13, t17 p3, p8, p9, p11, p18, p19S21 V 2
10ð6Þ t4, t7, t11, t15 t5, t8, t13, t17 p2, p3, p9, p11, p18, p19S22 V11 (9) t1, t9, t15, t18 t5, t8, t10, t17, t19 p6, p7, p8, p9, p11, p13, p16, p18, p19S22 V12 (9) t1, t4, t9, t15 t3, t5, t8, t10, t18 p6, p7, p9, p11, p13, p17, p18, p19S22 V13 (9) t1, t4, t15, t18 t3, t5, t9, t17, t19 p6, p7, p9, p11, p12, p16, p18, p19
See Tables 1 and 2 in [15] for S1–S18. S19 ¼ V3;f V4; p12; p17g , S20 ¼ V4;V1; p13; p18f g, S21 ¼ V8; p21; p22;f p24; p17; p10; p12g, and S22 ¼V6;V7; p2;p3;p10;
�p15;p20;p22;p23;p25p26g.
P13
t13
P12
t12
P11
t11
P10
t10
P9
P19t9 t8
P7
t7
P17
P6
P16
P8 P1P5
P4P3
P15
t2 t3
P2
t4
t6
R2
P14
P18
t5
t14 t1
Fig. 2 Benchmark 2 from [5]
Int J Adv Manuf Technol (2012) 60:625–629 627
respectively. Now, M p17ð Þ ¼ M p19ð Þ ¼ M V3ð Þ ¼ 0. Thetoken in VS2 goes to p3 to make M(VS2)=0. The totalnumber of tokens trapped in [S] is 4 ¼ M0 CSð Þ ¼P
V 2S M0ðV Þ while M(S)=0. Siphon S10 ¼ p7; p11; p12;fp17; p19;VS4;VS3g (see Table 1 in [17]) is synthesized fromthe core subnet obtained by adding TP-handle HTP2[t11 p17t10 p19 t7 VS3] with resource place r=p19 upon c3.S10 S10½ � ¼ p2; p3; p4; p5; p6; p9; p10Þf gð Þ is emptiable for asimilar reason.
For S8, two monitors are added as explained below for thetwo FBM: MΨ1 ¼ p3 þ p6 þ p9 þ p10 and MΨ2 ¼ p3 þ p5þp9 þ p10, where Ψ1 ¼ p3; p6; p9; p10f g and Ψ2 ¼ p3; p5;fp9; p10g. Two monitors are added using the FBM method:
V1
S8 (with V 1S8
� � ¼ Ψ1, M0ðV Þ ¼ 4� 1 ¼ 3, �V 1S8 ¼ t5; t7;f
t11g, and V 1�S8 ¼ t2; t6; t9f g) and V 2
S8 (with V 2S8
� � ¼ Ψ2,
M0 V 2S8
� � ¼ 3, �V 2S8 ¼ t6; t11f g, and V 2�
S8 ¼ t2; t4; t9f g). Theyimplement the following inequalities:
MðΨ1Þ ¼ Mðp3Þ þMðp6Þ þMðp9Þ þMðp10Þ < 4 ¼ MmaxðΨ1Þð1Þ
and
MðΨ2Þ ¼ Mðp3Þ þMðp5Þ þMðp9Þ þMðp10Þ < 4 ¼ MmaxðΨ2Þð2Þ
Only one monitor (as explained below) is added to thelast emptiable siphon S10 with S10½ � ¼ p2; p3; p4; p5;fp6; p9; p10g even though it is a mixture siphon with anon-sharing place p16. The only possible sets of marked
operation places for the unmarked S10 are Ψ1 ¼ p3;fp4; p5; p9; p10g � S10½ �, Ψ2 ¼ p3; p4; p6; p9; p10f g � S10½ �,Ψ3 ¼ p2; p3; p6; p9; p10f g � S10½ �, a n d Ψ4 ¼ p2; p4; p6;fp9; p10g � S10½ � with markings MΨ1 ¼ p3 þ p4 þ p5þp9 þ p10, MΨ2 ¼ p3 þ p4 þ p6 þ p9 þ p10, MΨ3 ¼ p2 þp3 þ p6þ p9 þ p10, and MΨ4 ¼ p2 þ p4 þ p6 þ p9 þ p10,respectively; the number of tokens in Ψ1 or Ψ2 or Ψ3
is five.By theory, four monitors are needed; one for each of
above Ψ. However, only one monitor is needed as explainedin [1] by exploring marking inequalities at the expense ofextra steps. The following theorem helps to eliminate theseextra steps.
Theorem 1 Let S (resp. S′) be a mixture siphon with the setof marked operation places Ψ (resp. Ψ′) when S (resp. S′) isunmarked under Ma. VS (resp. VS′) is the monitor added to S(resp. S′) with M0 VSð Þ ¼ Mmax Ψð Þ � 1 resp:M0 VS0ð Þ ¼ðMmax Ψ 0½ �ð Þ � 1Þ and H VSð Þ ¼ Ψ resp:H VS0ð Þ ¼ Ψ 0ð Þ. Ψ⊂Ψ′and ∀M∈R(N, M0), such that M Ψ 0ð Þ ¼ Mmax Ψ 0ð Þ )M Ψð Þ ¼ Mmax Ψð Þ. Then VS′ is redundant.
Proof After adding VS, M Ψ 0ð Þ ¼ M Ψð Þþ M Ψ 0nΨð Þ <Mmax Ψð Þ � 1þM0 R Ψ 0nΨð Þð Þ ¼ Mmax Ψ 0ð Þ � 1; hence, S′is controlled.
Now apply the developed theory to the example inFig. 2. Set ΨS10 of marked operation places for theunmarked S10 may be a superset of those ΨS8 of S8 andnot emptiable by theorem 1 since M ΨS10ð Þ ¼ M ΨS8ð Þ þM p4ð Þ < Mnax ΨS8ð Þ þM0 p14ð Þ ¼ Mnax ΨS10ð Þ (by Eqs. 1and 2), and S10 cannot become unmarked. For instance,ΨS8 ¼ p3; p6; p9; p10f g and ΨS10 ¼ p2; p3; p6; p9; p10f g.
Let ΨS stand for Ψ of S. Then Ψ1S8 � Ψ1
S10, and thecondition in theorem 1 is satisfied; i.e., ∀M∈R(N, M0), such
t h a t M Ψ1S10
� � ¼ Mmax Ψ1S10
� � ) M Ψ1S8
� � ¼ Mmax Ψ1S8
� �.
Thus, the monitor for Ψ1S10 is redundant. Similar conclusion
P13
t13
P12
t12
P11
t11
P10
t10
P9
P19t9 t8
P7
t7
P17
P6VS3
P16
VS4
P8 P1P5
P4P3
P15
t2 t3
P2
t4
t6
R2
VS2
P14
P18
t5
(P22)
(P21)
(P20)
t14 t1
Fig. 3 Partial controlled model obtained at the first stage in [1, 5] forthe net in Fig. 1
Table 2 Siphon-based controlled model for the net in Fig. 2.
S V(M0) V �S
�VS [VS]
S2 VS2 (1) t2, t11 t5, t13 p3, p11, p12S3 VS3 (3) t4, t5, t9 t7, t11 p5, p6, p9, p10S4 VS4 (2) t1, t11 t4, t5, t13 p2, p3, p4, p11, p12S8 V 1
S8 3ð Þ t2, t6, t9 t5, t7, t11 p3, p6, p9, p10S8 V 2
S8 3ð Þ t2, t4, t9 t6, t11 p3, p5, p9, p10S10 V 1
S10 4ð Þ t1, t6, t9 t2, t4, t7, t11 p2, p4, p6, p9, p10
628 Int J Adv Manuf Technol (2012) 60:625–629
applies to Ψ2S10. Ψ3
S10 ¼ p2; p4; p6; p9; p10f g is the onlyone such that ΨS8⊄ΨS10 with marking MΨS10 ¼ p2þp4 þ p6 þ p9 þ p10; the number of tokens in ΨS10 is five.Add monitor VS10 VS
�¼� Ψ3� �
= Ψ3� �� ¼ t1; t6; t9f g�
and�VS ¼ Ψ3
� ��n� Ψ3� � ¼ t2; t4; t7; t11f g; S ¼ S10Þ. The final
controlled model is shown in Table 2, the same as that in[1, 5]. The live controlled model can reach maximallypermissive 205 states without weighted control arcs. Notethat Ψ3
S10 � S10½ � and is more fragmented and hence usesmore control arcs.
Furthermore, one can reduce the number of controlarcs by setting [V]=[S] since only MΨ3 for S10 isreachable. The rest of unmarked-siphon states (MΨ1,MΨ2, and MΨ4) have been prevented by the monitors forS8. This explains why the controlled model by Hwang etal. employs fewer arcs.
3 Conclusion
Huang et al. [1] propose novel crucial MTSIs (CMTSIs) toreduce the number of MTSIs and hence the computationburden. In addition, the number of control arcs is fewer.Experimental results show that the proposed control policyis the most efficient algorithm among the closely relatedapproaches while the resulting model remains maximallypermissive. However, Huang et al. offer no hints on why itemploys fewer control arcs. This paper thus develops a newtheory to uncover the secret.
References
1. Huang YS, Pan YL (2010) Enhancement of an efficient liveness-enforcing supervisor for flexible manufacture systems. Int J AdvManuf Technol 48(5–8):725–737
2. Uzam M (2014) The use of Petri net reduction approach for anoptimal deadlock prevention policy for flexible manufacturingsystems. Int J Adv Manuf Technol 23(3–4):204–219
3. Uzam M (2002) An optimal deadlock prevention policy for flexiblemanufacturing systems using Petri net models with resources and thetheory of regions. Int J Adv Manuf Technol 19(3):192–208
4. Uzam M, Zhou MC (2007) An iterative synthesis approach to Petrinet based deadlock prevention policy for flexible manufacturingsystems. IEEE Trans Syst Man Cyber A 37:362–371
5. Li ZW, ZhouMC, JengMD (2008) A maximally permissive deadlockprevention policy for FMS based on Petri net siphon control and thetheory of regions. IEEE Trans Auto Sci Eng 5(1):182–188
6. Li ZW, Zhou MC (2004) Elementary siphons of Petri nets andtheir application to deadlock prevention in flexible manufacturingsystem. IEEE Trans Syst Man Cyber A 34:38–51
7. Piroddi L, Cossalter M, Fumagalli I (2009) A resource decouplingapproach for deadlock prevention in FMS. Int J Adv ManufTechnol 40(1–2):157–170
8. Chao DY (2009) MIP iteration-reductions for deadlock preventionof flexible manufacturing systems. Int J Adv Manuf Technol 41(3):343–346
9. Chao DY (2009) Direct minimal empty siphon computation usingMIP. Int J Adv Manuf Technol 45(3–4):397–405
10. Uzam M, Li ZW, Zhou MC (2007) Identification and eliminationof redundant control places in Petri net based liveness enforcingsupervisors of FMS. Int J Adv Manuf Technol 35:150–168
11. Abdallah IB, EIMaraghy HA (1998) Deadlock prevention andavoidance in FMS: a Petri net based approach. Int J Adv ManufTechnol 14(10):704–715
12. Huang YS (2007) Design of deadlock prevention supervisors forFMS using Petri nets. Int J Adv Manuf Technol 35(3–4):349–362
13. Li ZW, Wei N (2007) Deadlock control of flexible manufacturingsystems via invariant-controlled elementary siphons of Petri nets.Int J Adv Manuf Technol 33(1–2):24–35
14. Uzam M, Wonham WM (2006) A hybrid approach to supervisorycontrol of discrete event systems coupling RW supervisors to Petrinets. Int J Adv Manuf Technol 28(7–8):747–760
15. Chao DY (2006) Computation of elementary siphons in Petri netsfor deadlock control. Comp J 49(4):470–479
16. Chao DY, Liu GJ (2010) A simple suboptimal siphon-based controlmodel of a well-known S3PR. Asian J Cont. doi:10.1002/asjc.29
17. Chao DY (2011) Improvement of suboptimal siphon- and FBM-based control model of a well-known S3PR. IEEE Trans Auto SciEng 8(2):404–411
Int J Adv Manuf Technol (2012) 60:625–629 629
Recommended