This discussion paper presents the design and implementation of a decision-making framework for simulating production planning in manufacturing. The framework integrates techniques from Process Mining, Business Process Simulation, and Visual Analytics to analyze and interpret historical production event data. The business case centers on a manufacturing company specializing in sanitaryware production. The results highlight the potential of the framework to automatically generate production planning simulations, improving decision-making support.
The era of Industry 4.0 (I4.0) is characterized by the availability of a large variety of
Internetof-Things (IoT) devices that monitor the evolution of several real-world objects of interest and
produce a massive amount of data and events [
Production planning defines how products are manufactured, detailing production processes,
dependencies, human and technological resources, and schedules to meet customers’ demands
(e.g., on-time delivery) and organizational needs (e.g., reduced process cycle time). Production
planning is mainly performed with the support of Manufacturing Execution Systems (MESs),
Product Lifecycle Management Systems (PLMSs), and Discrete Event Simulation tools (DESs).
However, these solutions require a relevant manual efort performed by human experts to specify
the potential production plans of a company. They typically rely on digital models developed
during the design phase, thus representing outdated situations, and model obsolescence is a
critical issue because it might lead to wrong decisions on production that can not be accurately
predicted [
To maximize their potential, production planning tools must take advantage of the Big Data
from production processes by leveraging I4.0 technologies and adjusting automatically to the
dynamic changes in production [
To mitigate this issue, in this discussion paper, we present a framework that integrates some techniques from the I4.0 realm, namely Process Mining, Business Process (BP) Simulation, and Visual Analytics, to address two research challenges for enhancing decision-making support in production planning, namely: • (C1) the automated generation of up-to-date digital models of the production process from raw production event data; • (C2) the elaboration of digital models for extracting insight-ready data to improve the quality of decision-making activities.
Specifically, Process Mining techniques [
In this paper, we summarize the key concepts discussed in [
The rest of the paper is organized as follows. Section 2 describes the CER business case. Section 3 details the design of the proposed framework and its implementation in the context of CER. Finally, Section 4 concludes the paper by discussing lessons learned and future works.
CER is a leading Italian company in the sanitaryware industry that targets innovating ceramic production by adopting advanced I4.0 automation technologies integrated with an IoT infrastructure to keep up with the ever-increasing demand for quality products.
CER experiences a continuous stream of purchase order requests, encompassing numerous ceramic-based sanitary items that need to be manufactured and delivered within a specific time frame. The CER production process relies on robotic arms to carry out the manufacturing phases. Specifically, starting from a CAD (Computer-Aided Design) 3D virtual prototype of a sanitary product, a casting phase generates its initial mould model, which is then manipulated by the subsequent steps of the production process (drying, finishing, glazing and firing) to obtain the final product as desired at design-time.
Given the high volume of daily requests, even if the company is equipped with some production lines to parallelize the manufacturing of diferent items, it can only fulfill a limited number of production orders each day. This limitation arises from the manual intervention of expert decision-makers required to schedule the production and delivery of new orders, considering the production queue, the history of previously approved orders, the stock availability, etc. This activity involves variable factors such as the time needed to set up the production process for a new order and the non-fixed duration for a robotic arm to handle a ceramic artifact depending on the specific item being produced.
In the following section, we show how the proposed framework was designed and implemented over the CER case to forecast the optimal timing for producing and delivering new orders automatically while meeting customer deadlines and minimizing the changes in the CER’s production process. This is achieved by simulating diferent production plans and evaluating their insights through visual support.
From a methodological perspective, the proposed framework has been conceptualized and designed for addressing the two research challenges (C1) and (C2) discussed in Section 1. The framework consists of 5 operational stages to be applied in sequence: (i) Event Log Extraction, (ii) Process Discovery and Parameters Estimation, (iii) Model Tuning and Context Management, (iv) BP Simulation, and (v) What-if Analysis, as shown in Figure 1.
Expert
Model Tuning and Context Management
Historical Production Data incoming producton data
Event Log Extraction
Simulated
Scenarios What-if Analysis Event Log Extraction. Starting from the historical production data stored in the company’s databases (DBs), e.g., those data keeping track of the evolution of production over the years, as well as incoming data related to new products to be manufactured, the first stage of the framework is to generate an event log that encapsulates all the information about the production phases involved in the end-to-end execution of the production process. Specifically, an event log is a collection of traces. A trace represents the execution of an instance of the production process. Each trace contains a sequence of events, each related to a step of the production process. Events are associated with an activity label, a timestamp, and a trace identifier. The IEEE standard for storing, exchanging and analysing event logs is XES3 (eXtensible Event Stream).
In the case of CER, the event log was obtained by analyzing the three core relational databases (DBs) of the company, namely the Production DB, Warehouse DB, and Orders DB, together with the reception of a new order made by a customer. An order consists of the desired product types (e.g., washbasin of type ’A’ and water closet of type ’X’), the number of items for each product type (e.g., 300 items of type ’A’ and 500 of type ’X’), and an expected deadline for delivery. With this input at hand, an event log reflecting the historical production data of the product types included in the order was generated. In a nutshell, each trace in the event log encapsulates the information about the manufacturing phases involved in the end-to-end execution of the production process for a single item belonging to an ordered product type (e.g., looking at the previous example, the event log contains only data related to products of type ’A’ or ’X’). The procedure to obtain the event log from the company’s databases followed traditional event log extraction techniques [17, 18]. Process Discovery and Parameters Estimation. Once the event log is obtained, wellestablished Process Mining techniques can be leveraged to reveal fact-based insights into how the production process transpires. Specifically, the objective of this stage is twofold: (i) learning the sequencing of production phases to reproduce the behaviour observed in the log; and then, (ii) extracting the relevant time and cost information associated with each production phase (e.g., the total and average time of the production process, processing time and cost of each phase, average arrival times between traces, waiting time, etc.), together with an estimation of the human and technological involvement (e.g., cost of the involved resources, human resource calendars, etc.). Thus, the output of this stage consists of defining a digital model of the production process, which includes two elements: (i) a Business Process (BP) model detailing the control flow of the production process to be emulated, and (ii) the estimated production parameters, that are passed to the next stage of the framework.
In the case of CER, we relied on the Inductive Miner algorithm implemented in PM4PY4 to discover the BP model of the CER production process. Of course, for diferent orders, the possibility exists that slightly diferent BP models are discovered from the log. Then, we translated the event log into a relational DB through the RXES schema [19]. We used it to feed a customized version of the Process Mining technique described in [20], which enabled us to extract the relevant time and cost information associated with each manufacturing phase of the historical production of a product type included in the order, with an estimation of the human and technological involvement. When no historical information about a product type was available in the company’s DBs (e.g., for novel products to be manufactured), we employed its default production parameters.
Model Tuning and Context Management. To build realistic simulation scenarios, the BP model can be customized by expert users to comply with the technological and physical limits of the current production process. This includes incorporating in the model those contextual information not directly observable in the execution of the production process (e.g., market demand, financial constraints, etc.), or removing complexities that may hinder the correct interpretation of the model developed. Through a proper customization of the production parameters computed in the previous stage, this stage enables us to generate many simulation scenarios required to perform BP Simulation.
In the case of CER, we customized the behaviour of the BP model to make it compliant with the specific constraints of CER. Indeed, while some steps of the production process of CER involve robotic arms and machineries that can manipulate ceramic items associated with any kind of sanitary product (e.g., drying, glazing, and firing), other steps such as casting and ifnishing are constrained by technological and physical limits that must be considered in the BP model to build a realistic simulation scenario. For example, concerning the casting step, the CER plant is equipped with machineries configured with the specific mould stamp of the product the company will produce in a certain period. Switching mould stamps requires not only the direct and time-consuming involvement of human resources but even interrupts the working of some lines of the production process temporarily. To mitigate this issue, CER tends to produce sanitary items of a certain type in batches before changing the stamps. Consequently, based on the diferent product types in an order, the machineries available for the casting step 4PM4PY is a suite of state-of-the-art Process Mining algorithms for Python: https://pm4py.fit.fraunhofer.de/ must be wisely assigned to the diferent product types. For example, if an order includes only products of type ’A’ or ’X’, we may customize the BP model allocating /2 casting machineries to ’A’ and the others to ’X’. This is reflected by modeling two distinct BP activities for the casting phase in parallel branches, namely (for example) “Casting_A” and “Casting_X ”, each one associated with /2 machineries. Another possible (more drastic) solution is to assign all the casting machineries to ’A’, meaning that products of type ’X’ can start to be produced only after all products of type ’A’ completed the casting phase. Additional valid combinations can be implemented, allowing the creation of diverse BP models to simulate order production using various strategies. The point is that the BP model reflecting the CER’s production process can be structured in many variants based on the nature of the order being received by combining the sequencing and the number of the casting and finishing phases. These constraints represent an example of contextual information not available in process execution data but owned by CER domain experts that must be incorporated in the BP model before simulating the scenarios. In the context of CER, this task was performed automatically by a dedicated software component that generates suitable BP model variants considering their combination as realized in the historical production of orders of similar kind.
BP Simulation. It is the pivotal aspect of the framework. While traditional production planning requires a consistent involvement of human decision-makers to manually compute an optimal production strategy based on historical data, we rely on BP Simulation techniques to mimic the execution of the production process for specific product types. This allows us to automatically estimate the efectiveness of several potentially valid production plans that would require much human efort to be obtained manually. Once the computation of the multiple instances of the BP Simulation is completed, the results of the simulated scenarios are aggregated and visually presented to the decision-makers.
In the case of CER, we emulated independently each simulation scenario using the engine
running behind BIMP.5 We opted to use BIMP due to its wide range of simulation features and
lightweight performance that align perfectly with our specific needs. The simulation of a single
scenario configuration was repeated a fixed number of times (from 3 to 5, depending on the size
and complexity of the order); this is a common practice to minimize the potential outliers that
may arise from random factors during BP Simulation [
What-if Analysis. It is an approach for estimating the impact of changes to a production process in terms of time and cost measures. This stage facilitates informed decision-making by providing visual insights into the potential consequences of adjustments to a production plan according to the simulated scenarios.
In the case of CER, this stage was implemented through a visual analytic tool (cf. Figure 2)
that ofers a comprehensive range of visual comparisons among production scenarios. These
comparisons focus on metrics such as total cost/time, and the total number of mould changes
needed on machines required to implement the chosen production plan. Particular emphasis was
put on the presentation of the top three scenarios for the aforementioned metrics. In general, the
best scenario for cost, time, and mould changes depends on the configuration of the scenarios
in terms of the quantity of production lines used to parallelize the manufacturing of diferent
items, the order in which products are realized and the number of mould changes performed.
Furthermore, this user interface provides the opportunity to analyze individual scenarios
and get an overview of their relevant metrics and KPIs achievement. From a technological
point, the development of the visual analytic tool followed a User-Centered Design6 approach,
leveraging the well-established front-end Web stack (HTML, CSS and JavaScript) to ensure that
the implementation process does not require any additional local installation of dependencies
or frameworks. This aspect holds great significance as it reduces CER’s efort to seamlessly
integrate the tool into their existing information system. A preliminary user study evaluating
the usability and efectiveness of the What-If analysis component is presented in [
In this paper, we presented the design and implementation of a context-aware framework to simulate the production planning of a manufacturing company. Specifically, by leveraging historical and incoming data of production, our framework enables predicting the optimal 6ISO 13407:1999 timing for fulfilling new requests and meeting customers’ time requirements based on diferent simulated scenarios. By relying on the capabilities of Process Mining, BP Simulation and Visual Analytics, we analyzed and extracted valuable insights from historical production data towards building a proactive approach to support decision-making for production planning, allowing for the simulation of diferent production scenarios before their execution. Implementing the proposed framework for CER has demonstrated its ability to transform production data into digital models to simulate the efects of potential changes in the production process (C1), and improve predictive order forecasts, enabling more accurate decision-making (C2).
Despite the positive outcomes and advancements achieved through the implementation of the framework, there are certain limitations to acknowledge: (i) the efectiveness of the framework heavily relies on the quality and availability of historical production data. Incomplete or inaccurate data may lead to suboptimal predictions and recommendations. Therefore, ensuring data quality and integrity is crucial for maximizing the benefits of the framework [21]; (ii) our framework enables production planning tools to become automatically “aware” of the contextual information recorded in process execution data, while further contextual information not strictly related to the execution of the production process must be incorporated (if needed) by expert users into the BP model during the Model Tuning and Context Management stage; (iii) the complexity and variability of production scenarios make it challenging to generate realistic simulation models employing mainstream approaches to BP Simulation, which treat resources as undiferentiated entities grouped into resource pools. We have shown an example of this issue in Section 3 and the ad-hoc solution we found for properly managing the casting phase in a BP model. However, this limitation calls for novel BP Simulation techniques to capture the full complexity of manufacturing operations.
As future work, we would like to assess not only the entire framework’s efectiveness, including the accuracy of simulation scenarios and their alignment with real-world outcomes but also a comparative analysis with respect to alternative tools or approaches to production planning providing concrete performance metrics or objective measures.
This work could play a key role in the context of AI-augmented Business Process Management Systems (ABPMSs), that is, an emerging class of process-aware information systems empowered by AI technology for autonomously unfolding and adapting their execution flow [ 22]. Indeed, we envision extending this work in two diferent directions. Firstly, since the prominence and versatility of Large Language Models (LLMs) [23, 24, 25] have reached unprecedented heights, an additional future work could be to embed LLMs into the What-if Analysis stage to assist decision-makers in the selection of the optimal simulated scenario. Secondly, we envision leveraging the concept of process framing [26] to predict new possible production scenarios that break the boundaries imposed by the current production constraints, thus deviating from what is expected and ensuring process resiliency [27] and adaptation [28].
Acknowledgments. This work is supported by the H2020 project DataCloud (Grant 101016835), the Sapienza project FOND-AIBPM, the PRIN 2022 project MOTOWN and the PNRR MUR project PE0000013-FAIR.
[16] D. Roman, R. Prodan, N. Nikolov, A. Soylu, M. Matskin, A. Marrella, D. Kimovski, B. Elvesaeter, A. Simonet-Boulogne, G. Ledakis, H. Song, F. Leotta, E. Kharlamov, Big Data Pipelines on the Computing Continuum: Tapping the Dark Data, Computer 55 (2022) 74–84. [17] V. Stein Dani, H. Leopold, J. M. E. van der Werf, X. Lu, I. Beerepoot, J. J. Koorn, H. A. Reijers, Towards Understanding the Role of the Human in Event Log Extraction, in: Business Process Management (BPM) 2021 Int. Workshops, Springer, 2022, pp. 86–98. [18] D. Benvenuti, L. Falleroni, A. Marrella, F. Perales, An Interactive Approach to Support Event Log Generation for Data Pipeline Discovery, in: 46th IEEE Annual Computers, Software, and Applications Conf., COMPSAC, 2022, pp. 1172–1177. [19] B. F. van Dongen, S. Shabani, Relational XES: Data Management for Process Mining., in:
CAiSE Forum, volume 2015, 2015, pp. 169–176. [20] M. Camargo, M. Dumas, O. González-Rojas, Learning Accurate BP Simulation Models from Event Logs via Automated Process Discovery and Deep Learning, in: 34th Int. Conf. on Adv.
Inf. Sys. Eng. (CAiSE), Springer, 2022, pp. 55–71. doi:10.1007/978-3-031-07472-1\_4. [21] A. H. Ter Hofstede, A. Koschmider, A. Marrella, R. Andrews, D. A. Fischer, S. Sadeghianasl, M. T. Wynn, M. Comuzzi, J. De Weerdt, K. Goel, et al., Process-data quality: the true frontier of process mining, ACM Journal of Data and Information Quality 15 (2023) 1–21. [22] M. Dumas, F. Fournier, L. Limonad, A. Marrella, M. Montali, J.-R. Rehse, R. Accorsi, D. Calvanese, G. De Giacomo, D. Fahland, et al., AI-Augmented Business Process Management Systems: A Research Manifesto, ACM Transactions on Management Information Systems 14 (2023) 1–19. [23] M. Vidgof, S. Bachhofner, J. Mendling, Large Language Models for Business Process Management: Opportunities and Challenges, in: Business Process Management Forum BPM 2023 Forum, volume 490 of Lecture Notes in Business Information Processing, Springer, 2023, pp. 107–123. [24] A. Casciani, M. L. Bernardi, M. Cimitile, A. Marrella, Conversational Systems for AIAugmented Business Process Management, in: International Conference on Research Challenges in Information Science, Springer, 2024, pp. 183–200. [25] M. L. Bernardi, A. Casciani, M. Cimitile, A. Marrella, Conversing with business processaware large language models: the BPLLM framework, Journal of Intelligent Information Systems 62 (2024) 1607–1629. [26] M. Montali, Constraints for Process Framing in AI-Augmented BPM, in: Int. Conf. on
Business Process Management (BPM), Springer, 2022, pp. 5–12. [27] A. Marrella, M. Mecella, B. Pernici, P. Plebani, A design-time data-centric maturity model for assessing resilience in multi-party business processes, Information Systems 86 (2019) 62–78. [28] A. Marrella, M. Mecella, Continuous planning for solving business process adaptivity, in: Enterprise, Business-Process and Information Systems Modeling, Springer, 2011, pp. 118–132.