Technical PaperDynamic production system identification for smart manufacturing systems
Introduction
Knowledge of process requirements, system capacities, and system reliability are the premises on which control policies are formulated. In dynamic manufacturing environments, engineering change to the product, the process, and the production equipment can cause these premises to be violated and thereby make control policies less effective. An accurate, up-to-date model of the production system is essential to production control, but a challenge to maintain.
Both the need for a production system model and the challenge of maintaining it are more intense in smart manufacturing settings. The need is more intense because a key goal of smart manufacturing is to automated decision making [1]. Decisions concerning sequencing [2], line balancing [3], [4], and production system engineering [5] are sensitive to changes in process requirements, system structure, capacities, and reliability expressed in production system models. The challenge is more intense because smart manufacturing can make manufacturing more agile [1], and the changes brought on by increased agility must be reflected in the production system model. Change in process requirements is commonplace in manufacturing environments where products are evolving rapidly. Changes in system structure, capacities, and reliability are less common; but control policies are affected as much by changes in these dimensions as they are by changes in product and process.
Dynamic production system identification is a methodology that develops and updates a production system model that can provide information essential to performance analysis and control. The methodology (1) identifies a model that, like traditional statistical system identification [6] responds to stimulus accurately, (2) identifies system components, their properties, and interconnection, (3) identifies normative process for multiple job types, and (4) continually updates the model.
The production system model is a process model. Machine-learning methods of process mining typically develop such models using an analysis of frequently occurring events described in system logs. These methods fall short of addressing the challenge of dynamic production system identification in three important respects: (1) Rather than frequently occurring events, it is the infrequent, exceptional events that typically provide insight into system capacities and reliability. (2) Production system behaviour, especially machine blocking and starvation, are well-understood phenomena; an analysis of cause and effects could be used to guide search to an accurate system model. (3) Process mining lacks inherent means to update the model as the modelled system changes.
The production system model describes processes associated with International Society of Automation (ISA) Level 3 control problems [7]. Our methodology infers the production-system structure and capacities specifically for use in line scheduling and balancing processes (see Fig. 1). In the methodology, genetic programming, default causal knowledge, and probabilistic classification of exceptional conditions are used to evolve a population of individuals each representing a candidate model. The fitness of an individual is assessed with respect to its ability to (1) reproduce the content of logs describing typical Supervisory Control and Data Acquisition (SCADA) events, (2) in comparative steady-state analyses, respond to perturbations in workstation capacity with plausible differences in buffer occupancy and state sojourn times, and (3) detect critical job-type distinctions (e.g. that one job type requires significantly more processing time at some workstation than does another job type).
The main contribution of this paper is a robust methodology for dynamic production system identification. The paper investigates the value of genetic programming (GP) of Petri nets (PNs) in meeting its goals. GP on PNs is intended to facilitate adaptation of the methodology to diverse production system architectures and logging scenarios. The paper provides novel methods to interpret logs, validate the model, and learn from exceptional events.
Section 2 of the paper describes related work. Section 3 presents a Petri net model, the Augmented Queueing Petri Net (AQPN) which provides the model of process used in GP evolution. Section 4 describes how exceptional conditions, causal validation, and model updating are handled. Section 5 describes a case study that uses the methodology. Section 6 concludes the paper with an assessment of the methodology's limitations and a discussion of future work.
Section snippets
Related work
Process mining [8], [9], and advanced system identification methods [10], [11] provide semi-automated means to produce process and system models for various purposes including process conformance (i.e., determining whether or not the actual process being practiced conforms to the normative process). Typically, these methods have the goal of capturing the most frequent process patterns and exhibiting robustness to noise [12].
van der Aalst et al. [13] describe a process mining algorithm known as
Dynamic production system identification
The goal of any process modelling effort is to produce models fit for purpose [17]. Knowledge of system capacities is essential to the purpose of modelling production scheduling. For complex system engineering generally, and production system engineering particularly, capturing the most frequent process patterns will not be sufficient to create such a model. There are three interrelated reasons for this. First, the behaviour of complex systems under unforeseen circumstances cannot be predicted
Exceptional messages
A probabilistic neural net (PNN) [31] is computed for each individual to distinguish exceptional messages (messages not modelled through execution of the individual's PN as described in the interpretation process above) from ordinary messages. The fidelity and certainty with which the PNN classifies exceptional messages provides the component of fitness associated with the individual's probabilistic behaviour.
A PNN is composed of three layers (see Fig. 4). The input layer has one node for each
Evaluation
This section illustrates use of the methodology under conditions typical of automotive underbody assembly [38]. Underbody assembly systems are asynchronous assembly lines comprised of many workcenters. Each workcenter may be comprised of several automated units (e.g. robots) that work in concert to perform a sequence of operations. The controllers of the individual robots are capable of issuing messages. Given the appropriate genetic operators, it may be possible to apply the methodology to
Conclusion
An accurate, up-to-date model of the production system is essential to production system control. Our proposed methodology addresses the three challenges to identifying and updating this model: (1) developing a method for inferring system structure from exceptional messages, (2) demonstrating that causal knowledge can be used to guide search to an accurate system model, and (3) showing that GP provides inherent means to update the model as the modelled system changes.
Using an expressive Petri
References (38)
- et al.
Balancing U-shaped assembly lines with resource dependent task times: a simulated annealing approach
J Manuf Syst
(2014) - et al.
Modeling and solving mixed-model assembly line balancing problem with setups. Part II: A multiple colony hybrid bees algorithm
J Manuf Syst
(2014) Perspectives on system identification
Annu Rev Control
(2010)- et al.
Discovering simulation models
Inf Syst
(2009) - et al.
Functorial models for Petri nets
Inf Comput
(2001) - et al.
An improved immigration memetic algorithm for solving the heterogeneous fixed fleet vehicle routing problem
Neurocomputing
(2015) - et al.
The smart factory: Responsive adaptive, connected manufacturing, Tech. rep.
(2017) Balancing and sequencing of assembly lines
(1999)- et al.
Production system engineering
(2009) IEC 62264-1 International Standard, Enterprise-control system integration – Part 1: Models and Terminology
(2003)