-In process mining, basic descriptive statistics over observed events of event logs or streams, projected onto a process model, are typically used for performance analysis. The so-called performance spectrum is used for the fine-grained description of process performance over time, additionally revealing phenomena related to the behavior of multiple cases together. The performance spectrum computed from traces aligned with a process model allows performance analysis of processes with concurrency. However, performance spectra are used to describe performance only along the case level, leaving performance analysis and monitoring of other process dimensions out of scope. This paper presents an approach and tool combining a synchronous proclet system with a performance spectrum for multi-dimensional performance real-time monitoring and post-mortem analysis. While the tool is a proof-of-concept implementation, designed for analysis of the control-flow, resource and queue dimensions of logistic processes, the presented concepts are general.
Performance analysis is an important element in process
management relying on precise knowledge about actual process
behavior and performance for enabling improvements [
To address these limitations, the so-called Performance
Spectrum (PS) was introduced in our prior work [
A tool integrating the PS with a Petri net model, capable of
handling concurrent processes, is presented in [
For example, a Material Handling System (MHS), such as an 1The source code, video and further documentation available on https:
airport Baggage Handling System (BHS), can be considered as //github.com/processmining-in-logistics/psm/tree/pqr
a network of queues and resources connected according to the
system physical layout [
We present a tool1 that allows an integrated analysis and visualization of performance problems along the dimensions of control-flow, queues, and resources for the domain of MHSs.
It is a proof-of-concept implementation, which can use a
realtime event stream from the provided BHS simulation model
or a real MHS. We address the problem of multi-dimensional
performance analysis using a synchronous proclet system [
We mitigate the problem of understanding a process model by aligning its layout with Material Flow Diagrams (MFDs), the Copyright © 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). (c) bs starta completea completea starta enqa:b deqa:b pid1 pid2
t4 visualization widely used in the MHS domain, and by hiding implementation details when needed. We support real-time monitoring of the system performance by (1) token animation over the domain specific system visualization, similar to token animation in most process mining tools, and (2) through the PS being updated in real-time to show how the PS-based monitoring allows faster and more accurate analysis. Finally, to illustrate our concept in action, we show how a multidimensional analysis with PSa allows identifying root causes of performance outliers.
The remainder is organized as follows. Section II introduces PQR-systems and PSa. Section III explains how the tool satisfies the requirements of this section, and Section IV discusses evaluation and concludes the paper.
(a) a (b) b
Time
Fig. 1. The MHS conveyor (a), corresponding PQR-system (b) and PS (c).
In MHSs, materials (e.g., bags in BHSs) are typically moved III. Multi-Dimensional Analysis and Monitoring by conveyors. In Fig. 1(a) the conveyor moves bags from This section is organized as follows. We first introduce the location a to location b, the system tracks bags at those main components of the tool. Then we explain how to interpret locations through sensors that detect the front and back sides performance spectra of P-, Q- and R-dimensions of a PQRof each bag. We model MHSs as a synchronous proclet system system, illustrate how we adapt the visualization of the process called a PQR-system, where the only P-proclet models the model to needs of domain experts, and consider pros and cons conveyor layout, Q-proclets model conveyors as queues, and of performance analysis and monitoring using classical token R-proclets model resources. The life-cycle transitions start and animation versus visualization of PSa. Finally, we formulate a complete model how the front and back sides of a bag pass a brief guideline for performance analysis using an environment location. A PQR-system example is shown in Fig. 1(b). The with the integrated PS and process model that implements the P-proclet (shown in red) shows the logical layout of the system requirements of Section I. of Fig. 1(a). The Q-proclet a : b (shown in blue) models bag transportation as a FIFO queue with the minimal waiting time A. Tool Components twQ. The R-proclets a, b (shown in green) model resources Our tool consists of the PQR-system, the PSM, and the handling bags at the conveyor beginning (location a) and end optional BHS simulation model of a BHS (Fig. 2). The PQR(location b) as a single server (R-proclet of the resource b system visualizes the process model and provides a GUI is not shown to save space). Each resource has the minimal for filtering segments of di erent dimensions visualized by service time tsR required for handling a case (bag), and the the PSM. The simulation model provides a GUI to control minimal waiting time twR for becoming ready for handling the simulation scenarios and their animation. It sends each segment next case. The channels shown by dashed lines define how occurrence to the PSM as a datagram to simulate real-time transition occurrences of di erent proclets synchronize. As a monitoring of remote systems. In practice, the data for the tool result, each event is a part of three traces: of a case, of a queue, can also be streamed from a real system. and of a resource participating in the same step.
The PS is computed for segments. A segment is a pair of B. Performance Spectra of Multiple Dimensions
activities (x; y) where y directly follows x. It describes a step As we discussed in Section I, PQR-systems describe P-,
from activity x to activity y, e.g., a bag moving from location Q- and R-dimensions. Events, generated by the simulation
x to y. Each case advancing from activity x to y generates an model, have attributes that carry identifiers for corresponding
occurrence of the segment (x; y), which is characterized by two case notions and transition labels, so the PSM can assemble
timestamps (tx; ty). In Fig. 1(c), the PS of the PQR-system of traces per case (bag), queue, and resource for computing and
Fig. 1(b) is shown. Its top segment (as; bs) contains two segment visualizing their PSa, while the PQR-system GUI allows to
occurrences of cases pid1 and pid2 with timestamps (t1; t2) choose segments to be shown.
and (t3; t4) respectively. Multiple segments can be composed Fig. 1(b) shows the PS corresponding to “normal” work,
into the detailed PS [
We conclude this resource did not cause the delay but had to wait longer for the next case. Similarly, resource b had to wait longer (occurrence o6) to start handling the delayed case (occurrence o5), but the handling itself was not delayed.
Further, we analyse the PS of queue a : b. Similarly to the resources, each queue defines exactly one trace. A segment of a queue represents an element (bag) waiting in the queue. In the PS of queue a : b, we see the longer waiting time for the delayed bag (occurrence o4), so the queue indeed caused this delay. That proves also a longer waiting time of b (o6). Further in Section III-E we will show how to identify which system queue or resource initially caused a performance outlier.
Our simulation model also uses it to show a current location of each bag in the system. Additionally, it highlights stopped (blocked) conveyors in red. The PSM draws and classifies segment occurrences as soon as the corresponding events are generated by the system, and automatically scrolls the view towards the latest segment occurrence. As for ongoing segments the end activities and timestamps are not observed yet, it estimates them only if there is no choice in the control flow, i.e., when the end activity is known, using the observed minimal duration of process segments.
Let us first consider two screenshots of animation in the time moments t1 and t2; t1 < t2 in Fig. 4(a,b). The first screenshot shows normal work (no stopped (red) conveyors). In the second screenshot, the bags are not moving, because the conveyors are stopped (red). While the states are clear, a painstaking analysis of dozens or even hundreds of frames in between is required to understand how and why the incident was developing between those time moments. For processes with many cases executing simultaneously, and with a wide range of average durations of di erent process steps, such an approach is hardly feasible. In contrast, the PS allows to see the performance dynamics of the system for a time interval of interest at once, speeding up the detection of outliers during performance analysis or monitoring. In Fig. 4(c) we instantly see the delay in bag handling in segments s1; s3 s9, and how it propagates through the segments over time. As soon as an outlier is detected, the PS analysis can reveal its root causes, as we show next.
models in process mining tools oriented on diverse users. To overcome them, we followed two directions: (1) making models similar to process representations the domain experts are used to, i.e., MFDs, and (2) hiding information that is irrelevant for the current analysis phase or already known to the analyst.
For that, we aligned the PQR-system layout with the layout of the system MFD. In our tool, an MFD is present by the simulation model visualization (Fig.2(a)). We repeat this layout in our P-proclet, thereby making it easily recognizable by everybody familiar with the MFD. Then, we assume that implementation details of the Q- and R-proclets are extremely useful for getting to know the PQR-system first, but verbose afterwards. So we provide options to hide various details and dimensions.
single GUI. The PQR-system panel allows adding and removing segments of the P-, Q- and R-dimensions to the PS in the PSM, and the PSM allows their ordering needed for a particular analysis. Additionally, the PSM allows navigation from a PS segment back to the corresponding place of the PQR-system.
For example, in Fig. 4(c) the PSM shows (1) process segments s2 s9 forming a route from the BHS check-in link IN4 toward unit y S 0 diverting bags to scanner S 1, (2) resource segments s10 and s11 representing the resource of merge unit IN37 preceding directly diverting unit y S 0, and (3) queue segment s12 representing the queue between those merge and diverting units. Here we provide a basic guideline for the analysis of performance deviations, focused on blocking (a D. Token Animation versus Performance Spectra delay in performing of a particular process step) and high load
As we discussed in Section I, token animation over a system (a higher number of cases for a process segment within a given model is widely used for performance analysis and monitoring. time window). The guideline consists of the following steps:
1. Spotting a deviation. For that, we explore the P-proclet PS visually, e.g., in Fig. 4(c) we see longer occurrences (blockings) in segments s1; s3 s9, shown in orange. We select one of them, segment s3, for the further analysis. 2. Identifying the starting segment. For that, we analyze the PS, starting from the selected segment, against the direction of the outlier propagation. Concretely, high load typically propagates forward along the control flow, and blocking propagates backward. We add each next segment, containing the same type of deviation around the time of the previously spotted one, to the chain of segments until the next segment does not have deviations. The last segment of this chain had caused the whole chain of outliers. In our example, we have the chain hs3; : : : ; s9i. s9 is the last segment where the blocking is still observed, so it caused the whole chain of blocking. 3. Analysing the other dimensions. For the starting segment, we analyze the PS of the Q- and R-dimensions. For blocking, a longer queue waiting time or a longer resource service time can cause the delay, while for high load a higher number of cases accumulated, for example, during a recent blocking can cause a load peak. In our example, we study the R- and Qdimensions for segment s9. Segment s11 (state busy) has a long pause, i.e., it did not handle cases during the deviation time. Segment s10 (state idle) has a single long occurrence, i.e., a longer waiting time before the next case. Segment s12 (queue) shows multiple delayed traces whose waiting times are longer than the waiting time of the surrounding traces, so we conclude this queue caused the delay in s9 and the whole chain of blocking as well. 4. Domain-specific explanation. Using domain knowledge, the analysis results can be explain in terms of the underlying process or system. In our example, conveyor IN37 y : y S10 could be blocked by a bag, or it could have a technical malfunction, causing longer waiting times for all cases (bags) located on it. In turn, the preceding conveyors were gradually stopped one by one as they could not hand over bags to the further (already stopped) conveyors. The red arrows in Fig. 4(c) shows the incident development. Note, after s5 blocking propagates via merge unit IN34 x in two directions: to s4 and s1.
The tool combining PSa and MFDs was successfully
evaluated at Vanderlande for several systems and use cases, still
in the o -line mode. We showed that such an integration allows
a shallow learning curve and faster performance analysis for
domain experts. We expect further evaluations of the combined
PQR-systems and PSa for both analysis and monitoring. Our
tool still lacks automatic detection of performance patterns [
The research leading to these results has received funding from Vanderlande in the project “Process Mining in Logistics”.