Vadim Denisov, Dirk Fahland, Wil van der Aalst Predictive … · Vadim Denisov, Dirk Fahland, Wil...
Transcript of Vadim Denisov, Dirk Fahland, Wil van der Aalst Predictive … · Vadim Denisov, Dirk Fahland, Wil...
Predictive Performance Monitoring of Material Handling Systems
Using the Performance Spectrum
Vadim Denisov, Dirk Fahland, Wil van der Aalst
Baggage Handling Systems
1
Check-InX-
Ray
Screening &identification
Finalsorting
Happy path
Re-Circulation Causes Significant Delays
2
Check-InX-
Ray
Screening &identification
Finalsorting
Longer path
Insufficient Number of Screening Machine Operators
3
Check-InX-
Ray
Screening &identification
Finalsorting
Unavailability of Outgoing Conveyors
4
Check-InX-
Ray
Screening &identification
Finalsorting
Predictive Performance Monitoring of Material Handling Processes
5
Check-InX-
Ray
Screening &identification
Finalsorting
System Behavior
Predictive Model
System Operators
Op.number,Re-circulation
Action
What is the required number of
operators in +x minutes from now?
What is the amount of re-
circulation in +x minutes from now?
What Are Relevant Features?
6
• Bag color: red
• Final destination: Amsterdam
• Dimensions: 550x450x250mm
• Weight: 15kg6
• Utilization of the Loop Sorter: 90%
• Utilization of conveyors around the Sorter: 60%
• Current load on the X-Ray Screening Machine: 15 bags/min.
• Number of bags in the system: 8000
Intra-case features Inter-case features
What is the load on the screening machines? What is the amount of re-circulation in +x minutes from now?
One bag = one case
Bag properties System properties
Performance Spectrum [1]
7
a
b
Sub-trace: (ei(a, ta), ei+1
(b, tb))
Bins:• cases started• cases stopped• cases pending
normal speed
2 times slower
3 times slower
very slow
Time
a
b
ta tb
Step
Step
[1] V. Denisov, D. Fahland, and W. M. P. van der Aalst, “Unbiased, finegrained description of processes performance from event data,” in Business Process Management, M. Weske, M. Montali, I. Weber, and J. vomBrocke, Eds. Cham: Springer International Publishing, 2018, pp. 139–157.
Development of Problem Instance 2: Re-Circulation because of Unavailability of a1-a3
9
Check-InX-
Ray
Screening &identification
a1
a2
a3
b1
b2
b3
c1
c2
c3
loop
normal speed 2 times slower 3 times slower very slow
X-Ray:a2
a2:b2
b2:c2
a3:b3
b3:c3
X-Ray:a1
a1:b1
a1:loop
Severity
Time
Extracting Features from Performance Spectrum
10
Historic spectrum
Target spectrum
NowPrediction
horizon
Time
X-Ray:a2
a2:b2
b2:c2
a3:b3
b3:c3
X-Ray:a1
a1:b1
a1:loop
8(Y)
2(LB)
5(DB)
vij= [8, 2, 5]
Multi-Channel Performance Spectrum
11
NowPrediction
horizon
Time
X-Ray:a2
a2:b2
b2:c2
a3:b3
b3:c3
X-Ray:a1
a1:b1
a1:loop
Channel 1: SpeedChannel 2: Screening resultChannel 3: Sortation status
X-Ray:a2X-Ray:a2
Multi-Channel Performance Spectrum
Multi-Channel Performance Spectrum
12
Channel 1: SpeedChannel 2: Screening resultChannel 3: Sortation status
vns=<v1,…,vk> vn
e=<v1,…,vk>… …
v1s=<v1,…,vk> v1
e=<v1,…,vk>Se
gmen
ts
Binss e
s ns 1
………
2
1
3
In the paper: Generic definition of the Multi-Channel Performance
Spectrum Generic method for Performance Spectrum -based
feature extraction
Training& test sets
Model training
ML Model
13
Flatten
Sliding Window over Performance Spectrum
Problem Instances and Datasets
14
Check-InX-
Ray
Screening &identification
a1
a3
b1
b2
b3
c1
c2
c3
loop
Problem Instance 1Load on the X-Ray Screening Machines
Problem Instance 2Extra re-circulation because of unavailability a1-a3
Property Simulation model [2]
Time period 7 daysActivities 25Cases per day 1600Events per day 19.000Events, total 134.000
[2] https://github.com/processmining-in-logistics/psm/releases/tag/1.1.6
a2Real BHS of a major European airport
4 month85025.000-50.0001.000.000-2.000.000148.000.000
Implementation: Performance Spectrum Miner [3] and Scripts on Top of PyTorch
15
[3] Denisov, V., Belkina, E., Fahland, D., van der Aalst, W.M.P.: The performance spectrum miner: Visual analytics for fine-grained performance analysis of processes. In: BPM 2018 Demos. CEUR Workshop Proceedings, vol. 2196, pp. 96–100. CEUR-WS.org (2018)
https://github.com/processmining-in-logistics/psm/tree/ppm
Computational environment:• a single-machine configuration• 40 CPUs• 6 GPUs• 400GB RAM• the Performance Spectrum Miner in ProM 6.9• the PyTorch Framework
http://www.promtools.org/doku.php?id=prom69
Models and Baselines
16
Baselines:• Naïve baseline: current load• Data-driven inter-case feature encoding (DDICFE) [4] [5]
[4] Data-driven inter-case feature encoding, A.Senderovich, C.D.Francescomarino, and F.M.Maggi, “From knowledge driven to data-driven inter-case feature encoding in predictive process monitoring,” Information Systems, 2019.[5] https://github.com/processmining-in-logistics/psm/tree/ppm/ppm/src/main/scala/org/processmining/scala/intercase
Models:• Logistic Regression (LR)• Feedforward Neural Network (FFNN)
Experimental Results
17
11%
Predicted spectrum
50%
Observed spectrum
Baseline spectrum
50%
50%
7%
-30-20-10
01020
Error (%)
A
A
A
B
B
C
C
C
D
E
E
30%Predicted spectrum
Observed spectrum
Baseline spectrum
30%
30%
Experiment Approach Model RMSE, % of maxSim. model PS LR 12.1
DDICFE FFNN 21.2Naïve baseline - 35.8
Real system PS LR 7.0DDICFE FFNN 7.6Naïve baseline - 11.0
Experiment Approach Model RMSE, % of maxReal system PS FFNN 2.4
Naïve baseline - 3.0
Problem Instance 1Load on the X-Ray Screening Machines
Problem Instance 2Extra re-circulation because of exits a1-a3 unavailability
Conclusion and Future Work
18
• Processes: cases influence each other• Performance Spectrum:
• modelling a variety of performance features over time• capturing inter-case dependencies
• Multi-Channel PS: formulating a large class of performance prediction problems as a regression problem
• Generic methodology of solving this problem as a ML task
• Feasibility on MHS apply to Business Processes• Feature selection: domain knowledge is required include a formal process model• Validation in practice is missing higher accuracy and longer prediction horizon
Thank you!